AI Mistake Fingerprints: How to Spot Generated Content in 2025

A marketing director shares a polished product video. A procurement lead circulates a supplier report. A board pack lands in your inbox with a well-structured executive summary. Each looks credible. Each may have been generated, in whole or in part, by an AI system — and each carries the faint but legible fingerprints of that origin. The challenge for UK organisations in 2025 is not simply whether AI-generated content exists in their workflows, but whether the people making decisions based on that content know how to read those fingerprints before they act on them.

The narrative around AI detection has largely stalled at a superficial level: look for robotic phrasing, check for disclaimer language, run it through a detector tool. But the more useful — and more durable — approach is to understand why certain artefacts persist across image, video, and text generation, and what those specific patterns reveal about the fundamental constraints of how these systems are trained. The tells are becoming subtler, but they are also becoming more systematic. That combination rewards those who know what to look for.

The Physics Problem: Visual Artefacts in AI-Generated Images

Image generators have improved dramatically in their ability to produce convincing faces, plausible scenes, and stylistically coherent compositions. Yet they continue to fail at something that any child with a basic understanding of the physical world manages instinctively: consistent light and shadow logic. Shadows in AI-generated images often behave as though cast from multiple conflicting light sources simultaneously — or from no coherent source at all. Hands remain a notorious weak point, not because generators haven't been trained on images of hands, but because the sheer variability of hand positions in training data makes it statistically easier for the model to produce a plausible-looking approximation than a geometrically accurate one. The result is fingers that merge, bend at impossible angles, or cast shadows that contradict the image's apparent light source.

Reflections present a similar class of problem. Windows, glasses, and polished surfaces in AI images frequently reflect scenes that bear no relationship to the environment depicted elsewhere in the image. This is not a bug that will be patched in the next model release — it reflects the fact that these systems learn statistical relationships between pixels rather than internalising a model of how light physically behaves. For decision-makers reviewing visual assets — whether in marketing materials, due diligence documentation, or press imagery shared by counterparties — developing an eye for light and shadow inconsistency is a practical and transferable skill. A quick check: trace the apparent light source and ask whether every shadow in the image is consistent with it.

The Uncanny Valley of Motion: Video Tells

AI-generated video has entered a phase where, viewed casually, it can pass for real footage. Viewed carefully, particularly with an understanding of where generators struggle, it fails in characteristic ways. Blink patterns are among the most reliable indicators. Human blinking is involuntary, irregular, and contextually modulated — we blink more when processing information, less when concentrating on a fixed point. AI-generated faces tend to blink at intervals that are either too regular, too infrequent, or conspicuously absent during moments where a real person would naturally blink. The effect is subtle but creates an unease that viewers often register before they can articulate why.

Hair and fine detail at object boundaries remain technically expensive to generate convincingly in motion. AI video often exhibits a softness or flickering at the edges of hair, particularly where it meets a complex or moving background. Teeth, similarly, can shift subtly between frames in ways that real dentition does not. Perhaps more consequentially for organisations, AI-generated video presenters tend to exhibit limited and repetitive gesture vocabularies. A real speaker's hand movements are semantically linked to their speech in complex ways; AI gestures are broadly plausible but disconnected from specific meaning. If a video of a supposed expert or executive is being used to validate a claim or establish credibility, these motion tells are worth scrutinising systematically rather than trusting a gut impression.

The Assistant Brain: Recurring Patterns in AI-Generated Text

Text generation has perhaps the most sophisticated body of emerging critique, and for good reason — it is the most prevalent form of AI content in professional settings. The specific pattern worth understanding is what might reasonably be called the 'assistant brain' cadence: a writing register that reflects the statistical centre of gravity of an enormous corpus of training data skewed heavily toward instructional, explanatory, and helpful text. AI-generated prose tends to open with broad contextualising statements before narrowing to the specific, to use transitional phrases that signal structure rather than thought ('It is important to note that…', 'This highlights the need for…'), and to resolve complexity with a balanced summary that avoids commitment to a clear position.

There is also a characteristic relationship with hedging language. Human experts hedge selectively and purposefully; AI systems hedge reflexively, inserting qualifications that make text technically defensible but analytically weak. A report that concludes 'there are both advantages and disadvantages to this approach, and organisations should consider their specific context' has likely had minimal human intellectual engagement with the problem. Related to this is the tendency toward what researchers have begun calling 'false completeness' — AI text covers the expected headings and anticipated sub-points of a topic comprehensively, but rarely surfaces the genuinely unexpected insight or the specific operational friction that a domain expert would flag. The text is shaped by what questions are commonly asked about a topic, not by what is actually true or important about it in a given context.

Why These Artefacts Persist: A Training Data Problem, Not a Bug

A common misconception is that the artefacts described above are essentially engineering problems — implementation details that will be resolved as models improve. The more accurate framing is that they are symptoms of a fundamental epistemological gap. Current generative models learn from the statistical relationships within their training data. They do not learn from first principles, do not maintain a causal model of the physical world, and do not have access to the experience of actually inhabiting a body, watching light move, or forming a considered opinion. The outputs they produce are, in a precise sense, very sophisticated averages — plausible recombinations of patterns rather than reasoned constructions.

This matters because it sets a realistic expectation for how detection will evolve. As generators improve, the artefacts will become less obvious but they will not disappear — they will migrate to higher-order properties of the output. The hand problem will improve; the light logic problem will partially improve; but the absence of genuinely surprising insight in text, or of the micro-inconsistencies that make real video human, will persist as long as the underlying architecture remains fundamentally statistical. Organisations investing in detection capability now should focus not on learning to spot today's tells but on developing the analytical habits — the instinct to ask 'what would a real expert have said here that this doesn't?' — that will remain valid as the surface presentation improves.

For senior leaders and technical leads, the practical implication is straightforward: build structured scepticism into the workflows where AI-generated content poses the greatest risk. This does not mean treating every document or image as suspect, but it does mean identifying the specific contexts — vendor materials, research summaries, commissioned reports, public-facing communications — where the consequences of undetected AI slop are material. In those contexts, establish review steps that go beyond surface plausibility. Ask for the source reasoning, not just the conclusion. Look at the light in the image. Notice whether the video presenter ever gestures in a way that is specific to the sentence they're delivering rather than generically emphatic.

Detection tools have a role, but they should be treated as a starting signal rather than a verdict — their false negative rates remain significant, and sophisticated actors are already using detection evasion as part of their workflow. The more durable investment is in the human capacity to read the fingerprints: to understand why specific classes of error persist, and to know what questions to ask of an output that is trying very hard to look like it came from a mind. That critical literacy is now a practical professional skill, and organisations that develop it systematically will be better positioned than those waiting for a technical solution to a problem that is, at its core, epistemological.