This post covers where the verification process stands right now: what metadata can and cannot tell you, what visual artifacts are still worth checking, why reverse image search behaves differently with AI content, and what the emerging detection tools can and cannot be trusted to do.

Subscribe now

Metadata Analysis

Metadata is often the first thing practitioners check and the first thing that gets stripped or altered, intentionally or not. Most images shared on social media have already lost their original EXIF data through platform compression and re-encoding, which means the absence of metadata tells you very little on its own. A real photograph shared through a messaging app and then re-uploaded to a forum will often arrive with no more metadata than an AI-generated image.

What metadata can sometimes provide is a positive signal in the other direction. Some AI generation tools embed identifying information, either through the C2PA Content Credentials standard, which a growing number of platforms and generation tools support, or through tool-specific metadata fields that survive at least the first round of sharing. If an image carries C2PA credentials indicating it was generated or edited by a specific AI tool, that is strong evidence. The catch is that this metadata is easy to strip, either deliberately or as a byproduct of how an image gets re-saved, screenshotted, or re-compressed during sharing. Its presence is meaningful. Its absence is not.

For video, container-level metadata can sometimes reveal the software used to create or render a file, which is occasionally useful when that software is itself an AI generation tool rather than standard editing software. This is inconsistent enough that it should be treated as a secondary check rather than a primary method.

Share

The practical takeaway is that metadata analysis is worth doing because it occasionally provides a clean answer, but a negative or empty result should not be read as evidence that the content is authentic. It just means the metadata did not help.

Visual Artifacts and Inconsistencies

Despite the overall improvement in generation quality, certain categories of artifact remain useful, particularly in video and in images with complex scenes.

Hands, teeth, and ears remain disproportionately difficult for image generators to render consistently, especially in scenes with unusual poses or partial occlusion. This is less reliable than it was, but in a meaningful percentage of generated images these details still show subtle errors: an extra joint, asymmetry that does not match how the body actually works, teeth that do not align with the jaw structure.

Text rendering within images has improved substantially but still produces errors under scrutiny, particularly with text on signage, labels, or background elements that were not the focus of the generation prompt. Letters that are almost but not quite correct, logos that are close approximations of real brands but subtly wrong, and text that becomes incoherent at the edges of the frame are all still common.

Reflections, shadows, and lighting consistency are some of the more durable signals because they require the generator to maintain physical consistency across the entire frame, which is harder than getting any single element right. Check whether shadows fall in directions consistent with the apparent light source, whether reflections in glass, water, or mirrors show the same scene that is visible directly, and whether lighting on a subject’s face matches the lighting on their surroundings.

For video specifically, temporal consistency is the area where artifacts are most likely to appear. Background elements that subtly shift or warp between frames, hair or fabric that moves in ways inconsistent with physics, and faces that show small inconsistencies frame to frame even when any single frame looks convincing. These artifacts are often invisible at normal playback speed and become apparent when stepping through frame by frame or slowing the video down significantly.

Edge artifacts around the boundaries of people or objects, particularly where a subject meets a complex background, can show telltale blurring, color bleeding, or unnatural smoothness that differs from the texture of the rest of the image.

None of these signals is individually conclusive. The practical approach is the same as with any other OSINT verification task: look for a cluster of small inconsistencies rather than expecting one obvious tell, and weigh the cumulative picture rather than dismissing a single anomaly or treating a single anomaly as proof.

Reverse Image Search Limitations

Reverse image search is one of the most reliable tools in the OSINT toolkit for verifying real photographs, and it behaves very differently with AI-generated content in ways that practitioners need to account for.

When a real photograph has been used before, reverse image search can often find the original context, the original date, or prior instances of the same image being used in a different setting. This is the core mechanism behind a large portion of image verification work.

AI-generated images generally will not return prior instances through reverse image search, for the simple reason that the image did not exist before it was generated. A practitioner running a reverse image search and getting no results might interpret that as the image being original and therefore more likely authentic. With AI-generated content, the opposite can be true: a lack of any prior instances is consistent with both an authentic original photograph and a newly generated image, and reverse image search alone cannot distinguish between those two explanations.

There is a secondary use case that remains valuable. If a reverse image search does return a result showing the same image, or a very similar one, associated with an AI image generation platform, a prompt-sharing community, or a stock AI-image marketplace, that is strong evidence of AI origin. The search is not finding the “real” version of the image. It is finding evidence of the image’s existence within AI-generation ecosystems.

The broader point is that reverse image search needs to be understood as a tool that answers a different question for AI-generated content than it does for real photographs, and practitioners need to adjust their interpretation of both positive and negative results accordingly.

Emerging AI Detection Tools

A growing category of tools is being built specifically to detect AI-generated content, and the practical reality is that these tools are improving quickly but remain inconsistent, particularly across different generation models.

Detection tools generally work by analyzing statistical patterns in the image data that are characteristic of generation models, looking for the kinds of artifacts that generation processes tend to leave behind even when they are invisible to the human eye. Some tools focus on specific generation models and perform well against those models while performing poorly against others. Detection accuracy also tends to degrade significantly once an image has been compressed, resized, or re-encoded, which describes the vast majority of images encountered in real investigations.

The practical guidance for using these tools is to treat their output as one data point among several rather than a verdict. A detection tool returning a high confidence score that an image is AI-generated is meaningful and worth weighing heavily. A detection tool returning a low confidence score is much weaker evidence of authenticity, because it may simply reflect a generation model the tool was not trained to detect, or an image that has been processed in ways that obscured the relevant signals.

For video, detection tools generally analyze either individual frames using image-based methods or look for temporal artifacts across frames. The same caveats apply, with the added complexity that video compression is typically more aggressive than image compression and can further degrade the signals these tools rely on.

The detection tooling landscape is moving quickly, and tools that perform well against current generation models may perform poorly against models released six months later. This is not a space where a practitioner can learn a tool once and consider the skill set complete. Staying current with what tools exist and what their actual performance looks like against current-generation AI media is part of the ongoing work.

Why This Matters for Disinformation Tradecraft

The disinformation investigations covered in the previous post on this publication rely on establishing the authenticity and origin of content as a foundational step. A narrative built around a fabricated image or video is a different investigative problem than a narrative built around a real image used in a misleading context, and getting that distinction right at the start shapes everything that follows.

As AI-generated media becomes more difficult to distinguish from real content through casual inspection, the verification step becomes both more important and more demanding. The practitioners who treat this as a checklist item, running one tool and accepting whatever it returns, are going to miss things. The practitioners who understand what each method can and cannot tell them, and who weigh multiple imperfect signals against each other, are the ones whose assessments will hold up.

This is not a problem that gets solved once. The generation tools keep improving, the detection tools keep adapting, and the gap between them keeps shifting. Building the habit of methodical, multi-method verification now is what keeps a practitioner’s tradecraft relevant as that gap continues to move.

Some links in this post may be affiliate links. If you purchase through them, I may receive a small commission at no additional cost to you. This helps keep the blog running and the content free.

Subscribe now

Share

Leave a comment

The OSINTion Academy

Calibrated Citizen Academy

Calibrated Citizen Blog

The Pre-Mortem Intelligence Blog

True Blue Tactical Blog

Disinformation is not a new problem, but the scale at which false narratives now spread and the infrastructure that supports their spread has changed significantly. What used to require state-level resources to execute can now be done by small groups or individuals with access to social media accounts, a little patience, and some familiarity with how platform algorithms amplify content. The investigative response to that shift has also matured, and open source techniques now form the backbone of how researchers, journalists, and analysts trace where false narratives come from, how they move, and who is coordinating their spread.

This is applied OSINT work. It is not abstract analysis of a trend. It is methodical collection and analysis against a specific target: the network and infrastructure behind a piece of disinformation, and the people operating it.

Starting with the Narrative

Disinformation investigations typically start with a claim. Something is circulating that is demonstrably false or significantly misleading, and the question is where it came from and how it got to where it is now.

Subscribe now

The first analytical task is to establish what the claim actually says, in its most precise form, before attempting to trace it. This sounds obvious and gets skipped constantly. A claim that gets paraphrased or summarized before it is traced will produce a different search trail than the actual language of the original claim. Disinformation often spreads in specific formulations, and those formulations are what you search against.

Once the claim is precisely documented, the tracing process begins. The goal is to find and date the earliest instance of the specific claim in the form it is currently circulating. This is not the same as finding the first time the underlying subject was discussed. It is finding the first documented instance of this specific false narrative in this specific form. The difference matters because disinformation often layers false framing onto real events, and confusing the origin of the event with the origin of the false narrative is a common analytical error.

Tools like reverse image search, Google’s verbatim search, and archive services including the Wayback Machine and archive.today are essential here. Platform search with precise date filtering helps establish the timeline. The goal is to build a chronological record of the claim’s documented appearances, working backward from the current circulation to the earliest traceable instance.

Identifying Coordinating Accounts

Once the timeline of the narrative is established, the next layer of investigation focuses on the accounts amplifying it. Organic spread and coordinated spread look different when you examine them at the account level, and the techniques for distinguishing between them are well-established.

Account age relative to the current activity is one of the first signals. An account created recently that is posting at high volume on a specific topic, or an account that was dormant for an extended period and then suddenly became active around a specific event, warrants closer examination. Neither pattern proves coordination, but both appear consistently in documented influence operations.

Behavioral clustering is more diagnostic. When multiple accounts post the same content within a narrow time window, use identical or near-identical language, follow the same accounts, or engage with each other’s content in ways that suggest coordination rather than independent discovery, the pattern is worth mapping. One account doing this is noise. A cluster of accounts doing it simultaneously is a signal.

Profile construction tells a story when examined carefully. Accounts built for inauthentic amplification tend to have specific characteristics: profile photos that do not withstand reverse image search, bios that are either generic or suspiciously well-constructed, follower-to-following ratios that suggest artificial inflation, and posting histories that show either suspicious uniformity or gaps inconsistent with organic account use. None of these characteristics is conclusive on its own, but a cluster of them on a cluster of accounts engaged in the same behavior pattern is meaningful.

Share

The tools available for this layer of investigation include network visualization platforms like Gephi and Maltego for mapping relationship patterns, Botometer and similar services for behavioral scoring, and manual examination of account histories through platform-native search and archived captures of profile states over time.

Mapping the Influence Network

Individual accounts are rarely the full picture. Disinformation campaigns typically involve a distribution architecture: a set of accounts or pages that originate content, a set of amplifiers that push it to wider audiences, and downstream accounts that pick it up organically after it has gained enough apparent momentum to look credible. Mapping that architecture is the third layer of the investigation.

The origination layer is often the most difficult to identify because sophisticated operations deliberately obscure the true origin. Content may be introduced through low-profile accounts or obscure pages before being picked up by larger amplifiers, making the amplifiers appear to be the original source. Tracing backward through the share and repost chain, using timestamps to establish the actual order of propagation, is the method for getting past that misdirection.

Infrastructure analysis adds depth to the account-level mapping. Websites that host the original content or that multiple accounts in the network link to can be investigated through WHOIS records, registration history, hosting patterns, and cross-referencing against known disinformation infrastructure databases maintained by organizations like the Stanford Internet Observatory and the Digital Forensic Research Lab. Accounts that share the same external links, the same website references, or the same contact information in their profiles are connected in ways that transcend platform-level relationship mapping.

Metadata analysis of images and documents shared across the network can surface additional connections. Images shared by nominally independent accounts that share the same creation metadata, the same editing software fingerprint, or the same geographic EXIF data were likely produced from the same source. Documents with shared authorship metadata or creation timestamps that cluster around the same narrow window suggest centralized production regardless of how the distribution was structured.

Financial infrastructure is an underexplored area of disinformation investigation in the open source space. Accounts and pages that monetize engagement, that run advertising, or that solicit donations leave traceable financial infrastructure that can connect nominally independent operations to common ownership or funding.

What This Looks Like in Practice

A completed disinformation investigation of this type produces several distinct outputs. A timeline of the narrative’s spread, from first documented instance to current circulation, with the key amplification events identified. A network map of the accounts and infrastructure involved, with the origination layer, amplification layer, and organic downstream spread distinguished from each other. An assessment of whether the pattern reflects coordinated inauthentic behavior or organic spread of a false claim, with the evidence base for that assessment documented. And an identification of any infrastructure, accounts, or coordination patterns that connect this specific instance to prior documented campaigns.

That last element is where the investigative work connects to the broader intelligence picture. Disinformation operations rarely operate in isolation. Accounts, websites, and coordination patterns that appear in one campaign often appear in others, and documenting those connections is what transforms a single-incident investigation into an understanding of the actors and infrastructure behind a sustained effort.

It is also worth noting what this type of investigation does not produce without additional collection. It can establish that coordination occurred. It can map the infrastructure and accounts involved. It can identify the timeline and amplification pattern. It cannot, on open source collection alone, definitively identify the human beings behind the operation or their institutional affiliations without additional evidence. Overstating what the open source evidence supports is one of the more common and consequential errors in this field, and the standard of evidence for attribution claims should be treated with the same rigor applied to any other high-stakes analytical judgment.

The Tradecraft Foundation

The techniques described in this post are applications of standard OSINT tradecraft to a specific target set. The timeline reconstruction is source reliability and corroboration analysis applied to a narrative chain. The account analysis is entity profiling applied to social media infrastructure. The network mapping is link analysis applied to a coordination problem. None of it requires specialized tools that are unavailable to a trained open source practitioner. The methodology is the same. The target is different.

What it does require is analytical discipline, patience with a methodical process, and careful documentation of findings in a form that can withstand scrutiny. The subject matter is contested and the stakes for accuracy are high. An investigation that incorrectly attributes a disinformation campaign causes real harm to the people or organizations it misidentifies. The same standards of evidence, confidence calibration, and sourcing discipline that apply to any intelligence product apply here with at least as much force, and probably more given the public visibility this work often receives.

The OSINT community is well-positioned to contribute meaningfully to this space. The tradecraft is there. The analytical frameworks developed for entity investigation, network mapping, source reliability, and finished intelligence production map directly onto the disinformation investigation workflow. The application is a choice about where to direct skills that are already developed and a discipline that is already mature enough to do the work rigorously.

Some links in this post may be affiliate links. If you purchase through them, I may receive a small commission at no additional cost to you. This helps keep the blog running and the content free.

Subscribe now

Share

Leave a comment

The OSINTion Academy

Calibrated Citizen Academy

Calibrated Citizen Blog

The Pre-Mortem Intelligence Blog

True Blue Tactical Blog

Most geolocation work in OSINT focuses on the built environment. Street signs, architecture, road markings, storefronts, utility infrastructure. These are the signals practitioners learn first and reach for most often. They are also signals that can be obscured, cropped out, or simply absent depending on the image and how carefully the subject managed their surroundings before the shutter closed.

The natural environment is less controllable. A person trying to hide where they are does not think to remove the flowering plant in the background or position themselves away from the snake basking on the rock nearby. Flora and fauna appear in images constantly and get ignored constantly, because most practitioners have not developed the habit of reading them as geographic data.

They are geographic data. Not precise coordinates, but range data that can dramatically narrow a search area and, in some cases, add a temporal dimension that other geolocation signals cannot provide.

The Core Concept

Every species has a geographic range determined by climate, elevation, soil type, water availability, and ecological relationships with other species. That range is documented, often in significant detail, through scientific literature, species databases, field guides, and citizen science platforms. When a species appears in an image, its documented range becomes a geographic constraint on where that image could have been taken.

Subscribe now

The precision of that constraint varies. Some species have ranges spanning multiple continents. Others are endemic to a single mountain range or river basin. The investigative value scales with specificity, but even a broad range eliminates large portions of the map and focuses the search in ways that other signals can then refine further.

Combined with other geolocation data, a species identification can be the piece that locks a location. On its own, it establishes where to look. That is often enough to meaningfully advance an investigation.

Flora as a Geographic Signal

Plants are ideal geolocation signals because they cannot move. A plant appearing in an image was growing in that location when the photo was taken. That is a stronger geographic anchor than many other signals in the frame.

The example of the Mojave Prickly Pear illustrates the method well. That species has a documented range concentrated in the Mojave Desert and surrounding regions. An image showing a missing person alongside a blooming specimen narrows the search to that range, and then the bloom window, April through July, narrows it further temporally. You are not looking at the whole Southwest. You are looking at specific desert regions during a specific season.

The same logic applies across plant categories. Tree species carry geographic information because their ranges are constrained by climate and elevation. Coastal redwoods do not grow in the Rockies. Saguaro cacti do not grow in the Pacific Northwest. Spanish moss grows where humidity and temperature support it, which eliminates most of North America. Certain wildflower species bloom in ranges so specific that a positive identification puts you within a county or two.

Invasive species add an interesting dimension. Their ranges are expanding and are actively tracked, which means their presence in an image can sometimes be dated as well as located if the expansion front is well-documented in that area.

Share

The practical workflow starts with isolating the plant from the background and running it through a species identification tool. iNaturalist, PlantNet, and Google Lens all perform plant identification with varying degrees of accuracy. A confident identification goes next into the species range maps, which are available through sources like the USDA Plants Database, the Global Biodiversity Information Facility, and iNaturalist’s own range data. Cross-reference the range with whatever other geographic signals are present in the image and the search area contracts.

Fauna as a Geographic Signal

Animals are more complex than plants for two reasons. They move, and their ranges can shift with season and environmental conditions. Neither of these complications eliminates them as geolocation tools. They just require a more layered analysis.

The Ridge-nosed Rattlesnake is a strong example because that species, Crotalus willardi, has one of the more specific ranges of any North American rattlesnake. It is found in isolated mountain ranges along the Arizona and New Mexico border with Mexico, including the Huachuca, Santa Rita, and Chiricahua ranges. An image containing a Ridge-nosed Rattlesnake is not just pointing at the American Southwest. It is pointing at a handful of specific sky island mountain ranges. That is a significant constraint.

The same principle applies across the animal kingdom with varying levels of precision. Certain bird species are regionally endemic and appear nowhere else. The California Condor’s range, even in its recovered state, remains geographically limited. The Gila Woodpecker is a desert species whose range tells you something specific about where the image was taken. Insects are often highly localized. The Miami Blue butterfly exists in a range so small that a confirmed sighting is essentially a location.

Mammals are generally less useful because larger mammals tend to have broader ranges, but exceptions exist. Marine mammals can constrain coastal location. Certain bat species are range-limited. Livestock breeds and their regional distribution patterns can provide soft geographic evidence in agricultural settings.

The identification workflow for fauna parallels the flora approach. Isolate the animal, identify the species using iNaturalist, Merlin Bird ID for avian subjects, or reverse image search with species-specific terms, then map the range. The field guide literature for most species documents not just range but habitat preference within that range, which adds another layer of constraint.

Adding the Temporal Dimension

This is where flora and fauna analysis separates itself from most other geolocation methods. Many species signals are not just geographic. They are seasonal.

Flowering plants bloom on documented schedules that vary by latitude, elevation, and annual weather patterns, but fall within known windows. A plant in bloom is not just telling you where the image was taken. It is telling you approximately when. A blooming Mojave Prickly Pear is an April-through-July image. A blooming ocotillo is typically a March-through-June image at lower elevations. Documented bloom windows, combined with geographic range, produce a space-time constraint that can be cross-referenced against known timelines in an investigation.

Foliage state adds another layer for deciduous trees and shrubs. A tree in full leaf narrows the window to the growing season for that species at that latitude and elevation. A tree in bare winter dormancy does the opposite. Neither is as precise as a bloom window, but both eliminate significant portions of the calendar when combined with species identification and geographic range.

Migratory birds operate on predictable seasonal schedules. A migratory species present in an image narrows the window to the months it passes through or resides in that region. The presence of a Ruby-throated Hummingbird in an image taken in the eastern United States narrows the window to roughly April through October in most of its range. Breeding plumage visible on a bird tightens that further.

Insect activity is similarly seasonal. Certain butterfly species fly only during specific weeks of the year. Cicada emergences are among the most precisely timed biological events documented and can narrow an image to within days in regions with known emergence schedules.

None of this produces a timestamp. It produces a range of possible times that can be combined with other temporal signals in the image to narrow the window.

Limits and Honest Accounting

Range maps are representations of known occurrences and are not perfectly complete. A species appearing outside its documented range is unusual but not impossible, particularly for mobile animals, species undergoing range expansion, or escaped and introduced individuals. A positive identification should be treated as strong evidence that constrains the search, not as proof of location.

Identification tools make errors. iNaturalist and similar platforms are good and improving, but confident misidentifications happen, particularly with species that look similar to each other. Any identification driving investigative conclusions should be cross-referenced against field guide descriptions and, where stakes are high, reviewed by someone with direct expertise in the relevant taxa.

Indoor plants and pets complicate the analysis in different ways. A houseplant or an ornamental species purchased from a nursery is not a reliable geographic signal because it was placed there rather than growing naturally. The same applies to pets, zoo animals, or captive specimens appearing in images. Context matters for determining whether a species signal represents a natural occurrence or a human-introduced one.

Within those limits, biogeographic analysis is a legitimate and underused tool in the OSINT practitioner’s collection methodology. The built environment gets obscured. The natural environment is harder to hide and rarely thought about by the people being investigated. Learning to read species as geographic data is a skill that pays off in the cases where everything else in the frame has been carefully controlled and the only thing the subject forgot about was the plant growing behind them or the animal that wandered into the shot.

Leave a comment

The OSINTion Academy

Calibrated Citizen Academy

Calibrated Citizen Blog

The Pre-Mortem Intelligence Blog

True Blue Tactical Blog

Research personas have been part of investigative and intelligence work for as long as both disciplines have existed. The practice predates OSINT as a named field. Journalists go undercover. Law enforcement runs confidential informants. Investigators build cover identities to access communities that would otherwise be closed to them. None of this is new.

What is new is that building a convincing research persona used to require significant time, skill, and often a network of support infrastructure. AI tools have changed that equation in ways the OSINT community is still working through. For practitioners who have never built a persona before, AI makes the entry point more accessible than it has ever been. For experienced practitioners who have been doing this manually for years, AI removes a significant portion of the grunt work and improves consistency in ways that matter operationally.

This post covers both.

Why Research Personas Exist

Before getting into the how, the why matters. A research persona is an identity constructed for the purpose of accessing information or communities that would not be accessible to the practitioner under their real identity. The most common use cases in OSINT work are access to closed or invite-only online communities, source development in environments where the researcher’s affiliation would compromise their ability to gather candid information, and platform-level investigations where the subject would recognize and avoid a known researcher.

Subscribe now

The ethical frame here is straightforward: research personas are tools for access, not tools for deception as an end in itself. The goal is to observe, collect, and report accurately. A persona used to infiltrate a community spreading disinformation about a public health crisis is serving a legitimate research function. The same persona used to spread that disinformation is not. The distinction is about what the persona does once it has access, and that distinction sits entirely with the practitioner.

It also bears stating that research persona use carries legal and platform terms of service considerations that vary by jurisdiction, platform, and context. Those are the practitioner’s responsibility to understand before deployment.

What Makes a Persona Fail

Before building, it is worth understanding why personas get burned. The most common failure modes are thin backstory, inconsistent voice, implausible history, and activity patterns that do not match the claimed identity.

Thin backstory means the persona has a name and maybe a profile photo but nothing underneath. Ask it a question about its hometown and it cannot answer. Reference a cultural touchstone from its claimed background and it does not respond naturally. Communities, particularly tight ones with high social awareness, probe new members. A persona that cannot pass basic social interrogation does not last.

Inconsistent voice is subtler but just as damaging. If a persona claims to be a 45-year-old tradesman from rural Georgia and writes like a 28-year-old marketing professional from Austin, people notice. Not always consciously. But they notice.

Implausible history means the account was created yesterday but claims years of involvement in a niche community, or it has a job title that does not match its stated experience, or details that contradict each other across different conversations. Experienced community members keep track of things new members say. Inconsistencies accumulate.

Activity patterns are the operational failure mode. A persona that only posts during a 9-to-5 window in one time zone, never engages on weekends, and shows no organic activity outside the specific topic being investigated looks like what it is.

Where AI Changes the Work

AI tools address the first three failure modes directly and, when used thoughtfully, the fourth.

Backstory generation is where AI earns its place fastest. A well-prompted language model can produce a detailed, internally consistent personal history: childhood location, family structure, formative experiences, work history, hobbies, opinions on topics adjacent to the investigation, cultural references appropriate to the claimed background, and speech patterns that match the demographic profile. What used to take hours of careful manual construction can be drafted in minutes and refined from there.

Share

The key word is refined. AI-generated backstory is a first draft, not a finished identity. The practitioner needs to read it, internalize it, identify anything that does not feel authentic, and revise until it holds up. The persona is not the document. The document is a reference for the practitioner operating the persona. If you have not internalized the backstory well enough to answer questions in real time without checking notes, the persona is not ready.

Voice consistency is another area where AI is genuinely useful. Once a backstory and demographic profile are established, an LLM can generate sample posts, comments, and conversational responses in a voice consistent with that identity. This serves two purposes. First, it gives the practitioner a corpus of reference material for how the persona sounds. Second, it can help draft content for the persona’s public-facing activity during the warm-up phase, when the account is building organic history before being deployed for the actual research purpose.

Backstory coherence checks are underused. Feed an LLM the persona document and ask it to probe for inconsistencies, implausibilities, or details that would stand out to someone familiar with the claimed background. This is a faster version of the red team exercise that experienced practitioners do manually, and it catches things a single author misses because they wrote everything and cannot read it fresh.

Building a Warm Account

A research persona deployed immediately after creation is a liability. Platforms, communities, and alert members all apply more scrutiny to new accounts. The warm-up phase is the period of organic activity that makes the account look like it has been living its life before the investigation started.

AI-assisted content generation during this phase requires discipline. The posts and comments need to be about things the persona actually cares about based on its backstory, not the investigation topic. A persona built to investigate an extremist community should not be posting about extremist topics during warm-up. It should be posting about whatever else its identity would care about: sports, local news, a hobby, a professional interest. The investigation topic comes later, after the account has established a pattern of genuine-looking activity.

Scheduling tools and deliberate variation in posting times address the activity pattern problem. Build a posting schedule that reflects the persona’s claimed life: less frequent during claimed work hours, active in the evenings, occasional weekend activity, gaps that correspond to things a person in that life would actually be doing.

What AI Cannot Do

AI can generate backstory. It cannot generate judgment. The practitioner operating the persona has to know when to engage and when to hold back, when a question is a probe and when it is just conversation, when the persona is in danger of being burned and when it is fine. None of that comes from the backstory document or the voice samples. It comes from experience and situational awareness that only develops through doing the work.

AI-generated profile photos are a specific vulnerability worth addressing. Synthetic portrait images have improved significantly, but they have detectable artifacts and they do not hold up to reverse image search in the same way a real photograph does. A stolen or AI-generated photo that gets flagged ends an investigation. Persona photo strategy is a separate discipline and a longer conversation, but the short version is that AI-generated headshots are not a safe default.

Operational security for the practitioner behind the persona is also outside the scope of what AI helps with. Device separation, account isolation, network hygiene, and the procedural discipline required to keep the persona and the practitioner’s real identity from bleeding into each other are human problems that require human solutions. A persona that is operationally airtight but tied to the investigator’s real device or IP address is not actually protected. AI makes the persona more convincing. It does nothing to protect the person running it.

Connecting It Back

The first post in this series covered AI as a tool in the OSINT workflow, where it helps and where it fails. Research persona construction sits squarely in the category of legitimate AI assistance: it accelerates work that practitioners were already doing, improves consistency in areas where inconsistency has historically caused failures, and lowers the barrier to entry for newer practitioners developing the skill for the first time.

The failure modes from that first post still apply here. AI-generated backstory can contain implausibilities or demographic errors that the practitioner does not catch. Voice samples can carry artifacts that feel slightly off to a community member even if they cannot explain why. The output needs to be reviewed, refined, and tested before it goes anywhere near an active investigation.

A well-built research persona is an analytical asset. AI makes building one faster and more consistent than it has ever been. The judgment about when to use it, how to use it, and how to protect both the persona and the practitioner remains entirely human.

Leave a comment

The OSINTion Academy

Calibrated Citizen Academy

Calibrated Citizen Blog

The Pre-Mortem Intelligence Blog

True Blue Tactical Blog

It is not. Reliability is contextual. It is time bound. It is topic-specific. And treating it as a fixed attribute of a source rather than a dynamic assessment of a source in a specific context is one of the more common ways OSINT work goes quietly wrong.

The Shorthand Problem

The community has developed a set of reliability shorthand that feels intuitive and gets passed down through training and informal mentorship. Official government websites are reliable. Verified social media accounts are reliable. Mainstream news outlets are reliable. Academic sources are reliable. Forums and anonymous accounts are not.

This shorthand is not useless. It is a starting point. The problem is that too many practitioners treat it as an endpoint, and that is where the errors come in.

A government website is reliable for what the government wants you to know. It is an unreliable source for what the government is concealing, misrepresenting, or getting wrong. A verified social media account confirms that a platform authenticated an identity at a specific point in time. It says nothing about whether the content on that account is accurate, unbiased, or current. A mainstream news outlet may be reliable on domestic political reporting and systematically unreliable on a technical subject its reporters do not understand. Academic sources can be outdated, methodologically flawed, or operating from a theoretical framework that distorts their conclusions.

Subscribe now

None of this means those sources are worthless. It means the reliability assessment has to be specific to the claim you are trying to evaluate, not applied to the source as a whole.

Reliability and Accuracy Are Not the Same Thing

This distinction does not get enough attention. A source can be completely reliable and still give you wrong information. Reliability describes the source’s track record, motivation, access, and consistency. Accuracy describes whether a specific piece of information is correct. They are related but they are not the same measurement.

A reliable source with no direct access to the thing they are reporting on will give you sincere but wrong information. A source with direct access and strong motivation to deceive will give you information that is precisely calibrated to mislead. Treating reliability as a proxy for accuracy collapses a two-part evaluation into one and creates blind spots in both directions.

The intelligence community has formalized this distinction for decades, separating the reliability of the source from the credibility of the specific information. Most OSINT workflows do not make this separation explicit, which means analysts are often making a combined judgment without realizing the two components can point in different directions.

Reliability Is Not Static

A source that was reliable six months ago is not automatically reliable today. Organizations change leadership and editorial direction. Individuals develop agendas, get compromised, or shift their access to information. Platforms change their verification standards and content policies. Websites get taken over, spoofed, or quietly updated to reflect new interests. Any of these changes can happen without announcement and without any visible signal in the source’s surface presentation.

Treating source reliability as a durable rating rather than a current assessment means you are working from a stale evaluation. This matters most in long-running investigations where a source that was solid at the beginning of the collection phase may no longer be what it was by the time you are writing the assessment. It also matters in any environment where adversarial actors have motivation to corrupt or impersonate sources you trust.

Share

The question is not just whether this source has been reliable in the past. The question is whether anything has changed that should affect how you weight it right now, in this investigation, on this specific claim.

The Corroboration Trap

The community talks a lot about corroboration as the solution to source reliability problems. If multiple independent sources say the same thing, the information is more reliable. That logic holds when the sources are genuinely independent.

It breaks down when sources are drawing from a common origin without knowing it. A claim that originates in one place, gets picked up by a second outlet without independent verification, gets cited by a third, and then referenced by a fourth looks like four-source corroboration. It is one source echoing through a chain of repetition. The OSINT community calls this the laundry cycle, and it is more common in the current information environment than most practitioners account for.

This problem has gotten worse as the pace of information sharing has increased. A claim can complete that cycle in hours now. By the time an analyst is pulling sources on a topic, the apparent corroboration may already be fully circular, with every reference tracing back to a single originating report that was never independently verified. The fact that a claim appears in ten places does not tell you how many original sources exist. That requires tracing the citation chain, which is slower and less satisfying than noting that multiple outlets covered the same story.

Corroboration requires verified independence. Two sources that both cite the same original reporting are not corroborating each other. They are restating. The work of confirming actual independence is harder than it sounds, and skipping it because the corroboration looks clean on the surface is how single-source errors get embedded into multi-source assessments.

Motivated Sources and the Reliability Illusion

Every source has a perspective. Most sources have interests. Some sources have active motivation to shape what you conclude. None of that automatically makes them unreliable, but it makes the reliability assessment more complicated than a simple rating suggests.

A company’s investor relations page is a motivated source. So is a government’s official statement on a contested event. So is an advocacy organization’s research report, a competitor’s analysis of a rival’s technology, and an individual’s account of a conflict they were personally involved in. These sources can contain accurate, useful information. They can also contain carefully selected truths designed to produce a specific impression.

The distinction between lying and selective disclosure is important here. A motivated source does not have to fabricate anything to mislead you. They just have to control what you see. The facts they present may be entirely accurate. What they are not presenting may be the thing that changes the picture entirely. An analyst who evaluates a motivated source only on the accuracy of what is stated, without asking what is missing, is doing half the job.

The analyst’s job is not to dismiss motivated sources. It is to understand what the motivation is, how it might be shaping the information, and what you would need to see from a less motivated source to confirm or challenge what the motivated source is telling you. A motivated source that aligns perfectly with what you expected to find should raise your suspicion, not your confidence.

What the Community Should Be Asking

The useful questions about source reliability are specific and contextual. Does this source have direct access to the information they are providing, or are they reporting secondhand? What is their track record specifically on this type of claim, in this subject area, in this time period? Do they have interests that would be served by a particular version of events? Has anything changed about this source since the last time it was evaluated? Are the other sources that appear to corroborate this genuinely independent, or are they drawing from the same well?

None of these questions have simple answers, and none of them can be answered by looking at a source’s domain name, follower count, or verification badge. The shorthand the community relies on is a starting point for the evaluation, not a substitute for it.

There is also a documentation obligation that follows from this. If source reliability is a contextual judgment rather than a fixed rating, then that judgment needs to be recorded. What did you evaluate, on what basis, and at what point in the collection process? If a source you relied on turns out to have been compromised, biased, or simply wrong about the specific claim you used it for, a documented reliability assessment shows your reasoning at the time. The absence of that documentation is another version of the accountability problem described in the previous post on the intelligence cycle: conclusions without a traceable chain of reasoning are conclusions you cannot defend.

Source reliability is a judgment call that has to be made freshly, specifically, and with the right questions in mind. The analysts who get it right are not the ones with the best shorthand. They are the ones who are most honest about what they do not yet know about a source, most rigorous about finding out, and disciplined enough to document the assessment so it can be reviewed when it matters.

Leave a comment

The OSINTion Academy

Calibrated Citizen Academy

Calibrated Citizen Blog

The Pre-Mortem Intelligence Blog

True Blue Tactical Blog

These are not writing problems in the general sense. They are failures of discipline-specific craft. Intelligence products are evaluated against a different standard than other professional documents because they are used differently. A report that does not meet those standards does not just read poorly. It fails to do the job it was built for, regardless of how strong the underlying analysis was. The consumer cannot act confidently on conclusions they cannot evaluate, and they cannot evaluate conclusions that are not presented with the sourcing and confidence framing that makes evaluation possible. That is a structural failure, and it happens before the reader gets past the first page.

The Introduction to Intelligence Report Writing course covers how to structure an intelligence report correctly from the beginning. You will learn how to write for different consumer audiences without losing analytical precision, how to surface key judgments where they will be found and understood, how to attribute sources at a level of specificity that lets a reader assess what they are working from, and how to convey confidence and uncertainty in language that communicates both clearly rather than obscuring either.

This course is for practitioners producing finished products who know the reports are not as strong as the underlying work. It is for analysts moving into roles where their writing will be reviewed by more experienced consumers or used to support significant decisions. And it is the natural follow-on to AITI: that course builds the analytical framework, and IIRW builds the communication layer that makes the analysis usable by the people it is meant to serve.

Taking both courses in sequence gives you the full development track that most open source intelligence training leaves out entirely. You come out with both the analytical foundation and the ability to express it in a format that holds up under scrutiny and actually serves the consumer. The course is $59 at https://academy.theosintion.com/l/iirw.

Subscribe now

Share

The OSINTion Academy

Calibrated Citizen Academy

The Pre-Mortem Intelligence Blog

AITI covers the intelligence cycle as a working framework, not a diagram to memorize. You learn where OSINT sits within the broader collection environment alongside other disciplines, and what that positioning means for how you scope and document your work. The course covers structured analytical techniques that reduce the role of unchecked intuition and the cognitive bias that accumulates when you are working close to a subject over time. And it covers what it actually means to produce a finished intelligence product rather than a collection summary, which is a distinction most open source training never makes explicit.

That last point is the core distinction the course is built around. A collection summary tells the reader what exists. An intelligence product tells the reader what it means, how confident you are in that interpretation, what the gaps are, and what the implications are going forward. Most OSINT practitioners spend their careers producing the first thing while being evaluated on whether they delivered the second. That gap does not close on its own through more collection experience. It closes through deliberate analytical development, which is what this course provides.

AITI is for practitioners who are already doing OSINT work and know their outputs do not always land with the weight the underlying work deserves. It is for people newer to the field who want to build analytical discipline from the beginning rather than retrofit it onto years of collection habits. And it is for anyone whose background is entirely in collection and who has never had formal exposure to the tradecraft that sits on top of it, whether that gap came from self-teaching, tool-focused training, or a community that never clearly separated collection from analysis.

If the posts on this publication about the intelligence cycle and source reliability resonated with you, AITI is where those concepts get built into a practical working skill set rather than remaining useful ideas without a structured application.

The course is $59 at https://academy.theosintion.com/l/aiti

Subscribe now

Share

The OSINTion Academy

Calibrated Citizen Academy

The Pre-Mortem Intelligence Blog

That approach works until it does not. An investigation that was not built inside a proper workspace is harder to hand off, harder to revisit, and harder to document when someone asks you to show your work. A collection of module outputs without a structured record of how they were generated is not an investigation. It is a set of results that cannot be fully explained or reproduced, which matters in any context where the work is going to be reviewed by someone else or used to support a significant conclusion.

The power of Recon-ng is not in any individual module. It is in the ability to build a structured, reproducible investigation that can be extended, audited, and reported from without starting over each time. Getting there requires understanding how workspaces function, how data flows through the tool, and how to make intentional choices about which modules to run, in what order, and why rather than running everything available and sorting through whatever comes back.

The OSINT Using Recon-ng course covers the tool the way it is meant to be used. Workspaces and how to structure them from the start of an investigation. Module selection and how to think about what you are asking each module to do. Data management inside the tool and how to keep a clean record of what was collected and when. And how to integrate Recon-ng into a broader investigation workflow rather than treating it as a standalone lookup utility that gets opened, used once, and closed.

This course is for practitioners who have used Recon-ng and know they are not getting everything out of it, and for people who have avoided it because the learning curve felt steep without a clear starting point. Both groups finish with a working methodology for using the tool consistently and a structure for building investigations that hold up under review. At $49 it is a reasonable investment for a tool that earns it back on the first serious investigation it properly supports.

The course is at https://academy.theosintion.com/l/osint-using-recon-ng.

Subscribe now

Share

The OSINTion Academy

Calibrated Citizen Academy

The Pre-Mortem Intelligence Blog

Most practitioners who have been doing this for a while have a set of go-to tools they trust. What they often lack is a structured framework for using those tools together, in a sequence that accounts for platform coverage, the way different tools index and return data, and the investigative logic that connects one platform pivot to the next. Without that framework, username investigations are as good as the practitioner’s memory of which tools exist and what order they tried them in last time. Results vary. Gaps go unnoticed.

There is also a documentation problem. Username investigation that is not recorded in a reproducible way is hard to hand off, hard to revisit, and hard to explain when someone asks you to walk them through how you got from the original identifier to the collection of accounts you surfaced. Process matters for exactly the reason it always matters in intelligence work: it is what makes a result defensible.

The OSINT Username Tools course covers the tools and the workflow together. Not a list of websites to visit, but a practical framework for username investigation that treats account discovery as an analytical process rather than a lookup exercise. You will learn how to approach platform coverage systematically, how different tools handle username searches and where their results diverge, and how to build an investigation that does not leave gaps because you did not know which tool to reach for or when.

This course is built for practitioners who are already doing username work and want to do it more consistently, and for people newer to OSINT who want to build the habit correctly from the start rather than accumulating a disorganized collection of bookmarks. Both groups finish with a clearer, more repeatable process for a skill set that shows up in nearly every open source investigation regardless of subject.

The course is $20 at https://academy.theosintion.com/l/osint-username-tools.

Subscribe now

Share

The OSINTion Academy

The Pre-Mortem Intelligence Blog

Open source intelligence has spent the better part of the last decade building out an impressive collection infrastructure. Tools, techniques, communities, certifications, conference talks, YouTube tutorials. Nearly all of it is aimed at finding information. Searching better, pivoting faster, building link charts, geolocating images, pulling records, tracing infrastructure. The collection phase of the intelligence cycle has never been better served by community resources than it is right now.

The rest of the intelligence cycle is a different story.

This is not a knock on collection. Collection is real work and it matters. But if your open source intelligence process starts with collection and ends with collection, you have not produced intelligence. You have produced research. That distinction is not semantic. It determines whether your outputs get acted on or filed, whether your clients get clarity or just more data, and whether you are developing as a practitioner or running the same loop at increasing speed.

The Intelligence Cycle Is Not Optional

The intelligence cycle runs through planning and direction, collection, processing and exploitation, analysis and production, and dissemination. Every phase feeds the next. The cycle does not begin with collection, and it definitely does not end there.

Planning and direction is where you establish what question you are actually trying to answer. It sounds obvious. It is skipped constantly. Practitioners dive into collection without a clearly defined requirement, pull everything they can find on a subject, and deliver a volume of information that the recipient now has to make sense of on their own. The pile is impressive. It is not intelligence.

Processing and exploitation is where raw collected data gets transformed into something analyzable. Translating foreign language content, extracting structured data from unstructured sources, deconflicting information from multiple collection streams. This phase is where a lot of OSINT workflows quietly collapse, because the practitioner moves directly from collection to a document without ever working through the material systematically.

Subscribe now

Analysis and production is where the intelligence actually gets made. This is the application of structured analytical thinking to your processed collection in order to produce an assessment that answers the original requirement. It involves weighing source reliability, applying structured analytic techniques to reduce bias, calibrating confidence, accounting for gaps, and generating a judgment rather than just reporting facts. This is also the phase most OSINT practitioners have received the least formal training on.

Dissemination is where the finished product reaches the decision maker in a format they can use. Not a raw file dump. Not an email with attachments. A product structured to answer the question that was asked, presented in a way that lets the consumer understand what you assessed, how confident you are, and what they should consider doing about it.

Most open source intelligence workflows hit collection hard and touch the rest lightly if at all.

The Difference Between a Finding and a Product

A finding tells you something exists. An intelligence product tells you what it means, how confident you are in that meaning, what gaps remain, and what it implies going forward.

In practical terms: a finding reads as “the subject maintains three active social media profiles under different names, two of which were created within the past 60 days.” That is solid collection. A practitioner who surfaces it has done real work.

An intelligence product takes that finding somewhere. Why create two new profiles in 60 days? What do the content patterns suggest about intent or affiliation? What does the timing align with externally? What would change your assessment? What does not fit, and why? What are the competing explanations and how do you weigh them against each other?

Share

The second version requires analytical judgment, structured reasoning, sourcing discipline, and a clear understanding of what the person asking the original question actually needs in order to act. That is a different skill set from finding the information, and the community has historically under invested in developing it.

The gap shows up wherever OSINT outputs feed real decisions. Security teams get research reports and have to determine threat relevance themselves. Corporate clients get background summaries and have to assess risk on their own. The collector has handed off the hardest part of the job to someone who was not there for the collection and does not have the full picture. That is not a failure of the individual practitioner. It is a failure of how the discipline has been developed and taught.

There is also a sourcing and accountability dimension here that does not get discussed enough. When a human analyst produces an assessment, there is a traceable chain behind it: sources consulted, reasoning applied, confidence weighed. When a practitioner delivers a collection summary without that analytical layer, there is no chain. There are findings. If one of them is wrong, or if context shifts the meaning of the data after the fact, there is nothing to review, nothing to audit, and no documented basis for the conclusions the recipient drew from the work. That is a professional exposure problem that tends to surface at the worst possible moment.

Why the Community Stays Stuck at Collection

Part of this is structural. Collection is visible and demonstrable in a way that analysis is not. Showing someone a link chart, a geolocation, a social network map, a spreadsheet of extracted records. All of it produces an immediate reaction. You can show them in a conference talk. You can build a tutorial around them. They are concrete.

Analysis is harder to package. Showing someone how you weighted competing hypotheses against each other, how you documented your confidence calibration, how you structured key judgments separate from supporting evidence. That work is less flashy and requires more baseline knowledge on the part of the audience to appreciate. The community gravitates toward what it can demonstrate, and what it can demonstrate is collection.

The training ecosystem reflects this. The certifications and courses that define the OSINT space are built almost entirely around collection methodology. Knowing which tools to use, which databases to query, which search operators to string together. Necessary skills. Not the complete skill set for producing finished intelligence.

Analytical tradecraft is its own body of knowledge. Structured analytic techniques, cognitive bias awareness, confidence standards, sourcing discipline at the assessment level rather than just the collection level. These are learnable skills that have been formalized in professional intelligence environments for decades. They do not appear in most OSINT certification tracks because those tracks were built by and for collectors. Courses like Accelerated Introduction to Intelligence (AITI) exist precisely because this layer is missing from most open source development tracks, and the parallel gap in how practitioners write and structure finished products is what something like Introduction to Intelligence Report Writing (IIRW) is designed to close. Both skills are teachable. Neither one comes from doing more collection.

What This Means for Your Practice

If you are newer to this space, here is the frame worth internalizing early: the collection phase is where you gather the material. The rest of the intelligence cycle is where you do the actual intelligence work. Getting fast and skilled at collection is valuable. Stopping there means you are delivering raw material to someone else and calling it a finished product.

If you are experienced and bristling at this framing, you probably already know what the analytical layer requires. You got there somehow, through training, mentorship, or hard-won experience. The question worth asking is whether the practitioners coming up behind you are getting that same development, or whether they are learning that collection is the whole job because that is what the community modeled for them. The answer to that question is largely determined by what the community chooses to teach and what it chooses to treat as optional.

The INT in OSINT is not decorative. The discipline has intelligence in its name because intelligence is supposed to be the output. Most of the community’s energy goes into getting better at collection, which is necessary and not sufficient. The cycle does not stop there, and neither should the practice.

Clients and organizations that commission open source intelligence work are increasingly sophisticated about what they are getting. They have started to notice when they receive research summaries and have to do the analytical work themselves. Practitioners who can deliver a finished intelligence product rather than a well-organized collection dump are going to occupy a different tier of this market. The ones who cannot are going to keep wondering why their work feels undervalued.

The answer is usually not that the collection was poor. It is that the cycle was incomplete. Every phase of the intelligence process exists because leaving it out costs something: clarity, accountability, utility, trust. Collection without analysis produces data. Analysis without structured reporting produces conclusions nobody can evaluate. The full cycle produces intelligence. That is the standard the discipline names itself after, and it is worth taking seriously.

Leave a comment

The OSINTion Academy

Calibrated Citizen Academy

Calibrated Citizen Blog

The Pre-Mortem Intelligence Blog

True Blue Tactical Blog

This post is aimed at both working practitioners and people newer to the OSINT space. If you are experienced, some of this will confirm what you already know. If you are just getting started, it might save you from learning a few hard lessons in front of a client or a deadline.

When I say AI here, I am mostly talking about large language models (LLMs) like Claude, ChatGPT, Gemini, and their API-connected cousins, as well as AI-assisted tools layered on top of them for specific tasks. Image analysis tools and AI-driven search layers fall into the conversation too, but LLMs are the center of gravity right now.

Where AI Actually Helps

Processing Large Volumes of Text

This is where AI earns its keep in OSINT work. If you have pulled a massive leak, scraped a forum, collected thousands of social media posts, or are working through a document dump, manually reading everything is not realistic. AI tools can summarize, extract relevant entities, identify recurring themes, and flag anomalies far faster than any human analyst working alone.

Subscribe now

The key framing here is triage, not conclusion. You are using AI to figure out where to look harder, not to tell you what the answer is. Feed an LLM a batch of forum posts and ask it to surface the usernames that appear most frequently in discussions around a specific topic. That is useful. Asking it to tell you who those people are in real life based on the same data is asking it to do something it is likely to fail at, often without telling you it failed.

Breaking Language Barriers

Historically, language gaps have been a serious constraint in OSINT. If you do not read Mandarin, Russian, Farsi, or Arabic, you are either dependent on human translators or you are ignoring a large portion of the open source environment. AI translation has gotten genuinely good, particularly for common languages, and LLMs can go beyond literal translation to provide cultural and contextual framing that a raw machine translation cannot.

This does not mean you take an LLM translation and treat it as verified. It means you now have a working draft to reason from, and for many investigations that is the difference between having a lead and having nothing. Unusual idioms, sarcasm, coded language, and slang are still areas where AI stumbles, and anything that matters should get reviewed by a human speaker or professional translator if the stakes are high enough.

Surfacing Patterns You Might Miss

There is a difference between an AI finding a pattern and an AI surfacing a potential pattern for you to investigate. The first is something you should be skeptical about. The second is genuinely useful.

When you are deep in a collection and trying to see the shape of something, cognitive fatigue is real. You start missing connections. AI tools are not tired. They can compare text across hundreds of documents and flag similarities in phrasing, structure, timing, or entity co-occurrence that a fatigued analyst might walk right past. The catch is that you still have to go verify every single one of those flags. AI is generating hypotheses, not findings.

Drafting and Summarization

Report writing is a real time sink in OSINT work, and it is often the part of the job that gets least attention because all the focus goes to collection and analysis. AI tools are solid at taking a set of verified findings and helping you structure them into a coherent narrative, draft executive summaries, or produce different versions of the same reporting for different audiences.

The critical rule here is that AI should only be drafting from your verified conclusions, not reaching conclusions of its own. You write the analysis. It helps you express it. That is the correct division of labor.

Brainstorming Pivot Points

Sometimes the hardest part of an investigation is figuring out what to look for next. You have a username, a photo, a company name, a partial email address. Where do you go from there? LLMs are surprisingly useful as brainstorming partners for this. Ask it what platforms typically allow a specific type of username format, what open databases might contain records for a type of entity, or what search combinations might surface a particular kind of information. It will not always give you good answers, but it will often give you ten ideas, and two of those will be things you had not considered.

This works because you are not asking it to retrieve information from the internet. You are asking it to reason about information structures and search strategies, which is something LLMs do reasonably well when the inputs are general.

Where AI Gets You in Trouble

Hallucination

This is the most dangerous failure mode in an OSINT context and it needs to be stated plainly. LLMs make things up. They do not always know they are doing it. They will fabricate citations, invent biographical details about real people, generate fake URLs that look real, produce plausible-sounding company histories that do not exist, and attribute quotes to individuals who never said them. The output looks exactly like real information. It reads confidently. There is no warning label.

In a research or casual context, hallucination is annoying. In an OSINT context, it is potentially catastrophic. An analyst who unknowingly builds a case on hallucinated information is not just wrong. They may be feeding a harmful false narrative about a real person, a real organization, or a real event. The downstream consequences of that can be severe.

The rule is simple but requires discipline: anything an AI tells you about a specific person, organization, location, date, or event needs to be independently verified against a primary source before it enters your analysis. No exceptions.

The Knowledge Cutoff Problem

LLMs are trained on data up to a specific point and they do not have live access to the internet unless a tool explicitly provides it, and even then that access is limited and inconsistent. If you are working a current event, tracking an active actor, or investigating something that has moved in the last few months, an LLM working from its training data alone is going to give you stale or incomplete information. More dangerously, it will often present that stale information with the same confidence as current information.

Share

AI-assisted tools that include web search can help here, but they introduce their own reliability questions. Know exactly what your tool is doing when it fetches external information, where it is pulling from, and whether the retrieved content is reliable.

Plausible and Wrong

This is a subtler version of hallucination and in some ways more dangerous because it is harder to catch. LLMs are very good at generating text that sounds like expert analysis. They synthesize patterns from their training data and produce output that is internally consistent and sounds authoritative. The problem is that plausible reasoning built on incorrect premises is still incorrect, and in OSINT that can mean an analytically coherent narrative that points entirely in the wrong direction.

Experienced analysts develop an instinct for when something feels too clean, too neat, too perfectly assembled. AI output tends to trigger that instinct less than it should, because the writing quality is high and the logic tracks on the surface. The habit you need to develop is verification first, regardless of how convincing the output is.

Embedded Bias

LLMs reflect the biases present in their training data, and that training data is not a neutral, balanced sample of human knowledge. Coverage of certain geographies, languages, communities, and topics is uneven. Some subjects are represented primarily through particular ideological or cultural lenses. Others are underrepresented entirely.

In OSINT, this matters because bias shapes what the model thinks is significant, what it treats as normal, and what it flags as anomalous. If your investigation touches on communities, regions, or subjects that are poorly represented or skewed in the training data, you need to be especially careful about letting AI drive your framing. You may be getting a distorted picture and not know it.

OPSEC and Data Leakage

This one does not get enough attention in the OSINT space. When you paste investigation data into a commercial LLM, you need to understand where that data goes. Most consumer-facing AI products have data handling policies that include using inputs for training, sharing with third parties under certain conditions, and storing conversations. Enterprise and API tiers typically have better protections, but you need to read the terms, not assume.

If you are working a sensitive investigation involving a real person, a client, classified or privileged information, or anything that has legal exposure, feeding that information to a commercial LLM without understanding the privacy implications is a real operational security risk. Use local models when possible for sensitive work. At minimum, understand what the tool does with your input before you give it anything that matters.

Attribution and the Chain of Custody Problem

OSINT has a documentation problem in general, but AI makes it worse. When a human analyst reaches a conclusion, there is a traceable chain: the sources consulted, the reasoning applied, the judgments made. When an AI generates a conclusion, that chain is opaque by design. You often cannot explain how it got there, which means you cannot defend it under scrutiny.

If your OSINT work ever ends up in front of a legal proceeding, a congressional inquiry, a law enforcement context, or even just a demanding client, “the AI said so” is not a defensible position. Any conclusion that enters your reporting needs to be traced to verifiable human-reviewed sources. AI output that has not been verified and documented cannot be part of that chain.

A Working Framework

Using AI in OSINT responsibly comes down to a few consistent habits.

• Use AI for acceleration, not for conclusions. Collection assistance, triage, summarization, translation drafts, and brainstorming are where it earns its place. Final analysis belongs to a human working from verified sources.

• Verify everything specific. Any name, date, URL, organization, quote, or biographical detail that came from or passed through an AI tool needs independent verification before it touches your reporting. Treat AI output the same way you treat a source you do not fully trust: interesting, potentially useful, and unconfirmed until proven otherwise.

• Document your methodology. Whatever role AI played in your process, note it. This protects you professionally and ensures that anyone reviewing your work understands where human judgment was applied and where it was not.

• Know your tool’s limits. Read the data handling policies. Understand the knowledge cutoff. Know whether the tool has internet access and what that access actually covers. Use local models for sensitive work when you have that option.

The analysts who use AI well are not the ones who trust it most. They are the ones who have a clear model of what it can and cannot do and apply it accordingly. That is not different from how good OSINT practitioners have always treated any other tool in the collection stack.

AI is not going to replace disciplined tradecraft. But ignoring it entirely means leaving real capability on the table. Get familiar with where it actually works, stay sharp about where it fails, and you will be ahead of most of the people currently either overclaiming or dismissing it.

Leave a comment

The OSINTion Academy

Calibrated Citizen Academy

Calibrated Citizen Blog

The Pre-Mortem Intelligence Blog

True Blue Tactical Blog

Read more

Background

As some of you may know, Spread Privacy, the parent company of the DuckDuckGo search engine, announced on 20 July 2021 two new services. The first is similar to the disposable email address…

Read more

Sign up now so you don’t miss the first issue.

Subscribe now

In the meantime, tell your friends!

Read more

I was tossing around ideas for a blog about Poshmark. My first inclination was to check Micah Hoffman’s WhatsMyName tool (also available as Profiler in Recon-ng and as a Web Application). The tool ha…

Read more

Read more

Read more

As the world continues to battle the COVID-19 pandemic, some outlets are sharing samples of malware, while others are trumpeting politically biased information from both sides of the proverbial aisle…

Read more

I have been collecting old computers from friends and family for some time, and I always said that I was going to stand up a lab. I have finally started to put …

Read more

After one week of research and trial and error, I have some more progress to publish/report. I started the week by installing the packages used in the books that I am reading. I wrote a simple shell …

Read more