What Is Your Evidence?

What Counts as Evidence in the Age of Artificial Intelligence?

Artificial intelligence does not eliminate flying elephants.
It helps them acquire a convincing explanation.

Fossa

From Storytelling to Documentation

This article continues the idea introduced in Flying Elephants Flying Elephants Why science requires more than an elegant hypothesis. The role of photographs, videos, specimens, and primary data in building and verifying … .

At first, knowledge was passed on as a story. Later came more formal descriptions: measurements, characters, Latin names, comparisons, diagnoses. Then biological illustration became a way to show the reader an organism they could not see for themselves.

But a drawing was still not the observation itself. It was the author’s reconstruction of an observation. When a drawing is the only record, it is fundamentally difficult to verify. Years later, we cannot know whether it represents a real observation, the author’s interpretation, or imagination.

Then came photographs and videos. They did not make every conclusion automatically true, but they did make independent verification more realistic. A photograph preserves a moment; a video preserves a process.

Not All Papers Are Documented Equally

The problem of weak evidence did not begin with artificial intelligence.

In fish pathology, for example, one often sees the following logic:

a smear or culture was taken -> a bacterium grew -> therefore it caused the disease The Smear Comes First. A reflection on why smear microscopy often matters more than bacterial culture in fish pathology. Real cases, logic, and the danger of … .

But there is a long distance between “a bacterium grew” and “this bacterium caused the death.” We need the clinical picture, photographs of lesions, tissue microscopy, inflammatory response, repeatability, exclusion of background microbiota, spatial association with the lesion, and in some cases experimental confirmation.

If those elements are missing, confidence in the conclusion should decrease. Not because the author is necessarily wrong, but because the reader is not given enough access to the original observation.

The absence of a photograph does not make a conclusion false.
It increases the share of trust placed in the author and reduces the share of independent verification.

When a Drawing Replaces a Photograph

For its time, biological illustration was a serious method of documentation. But in a paper where microphotographs, video, slide series, or preserved material are technically possible, a single drawing becomes a weak basis for an unusual conclusion.

In that situation the claim may be extraordinary, while the reader’s ability to verify it remains limited. This does not mean every such paper must be dismissed. But it is an ideal example of the problem: the more unusual the claim, the less it should depend on trust alone.

The more extraordinary the conclusion, the less it should rest on the author’s authority.

When the Photograph Became a Drawing Again

For a long time, scientific documentation developed in one direction: from a story toward material that could be shown to another person. A story could be retold. A drawing could be redrawn. But photographs and videos seemed closer to the event itself.

For nearly two centuries, each new documentation technology reduced the need to take the author’s word for it. Drawing gave way to photography; photography gave way to video. It seemed that the evidential foundation was becoming more reliable.

Generative AI has changed that familiar trajectory for the first time. Today it can create not only a story or a drawing, but also a photograph, a video, a voice, a table, a literature review, and a plausible explanation for an event that never happened.

A high-quality photograph or video no longer guarantees that the observation actually occurred. In this sense, we have unexpectedly returned to a situation familiar to researchers of the early twentieth century: the image once again requires trust in its source.

Back to the Future of Evidence

18th century "I saw it" Do you believe me?

19th century "Here is a drawing" Better

20th century "Here is a photograph" Excellent

21st century "Here is a video" Almost ideal

AI era "Here is a photograph" Where did it come from?

We have not returned to drawings. We have returned to the need to trust the author.

Paradoxically, technological progress has brought us close to the early twentieth century. Back then, a drawing required trust in the observer. Today, photographs and videos increasingly require the same kind of trust. We now sit with images and recordings much as earlier biologists sat with scientific drawings.

This does not make AI an enemy of science. It simply makes the origin of data part of the evidence.

The Main Point

At first, the weak point was the retelling. Then it was the author’s interpretation. Now the weak point has shifted toward the origin of the data.

Image quality no longer guarantees observation quality.
Image quality increasingly guarantees only image quality.

Text quality no longer guarantees evidence quality.
The quality of an explanation no longer guarantees that the object existed.

Artificial intelligence has not made the scientific method unnecessary. On the contrary, it has reminded us why the method is needed Scientific Method. A Practical Guide. How knowledge is formed, why confidence often outpaces understanding, and how the scientific method helps distinguish observation from … .

In the past, it was often enough to ask: “Is there a photograph?” Now we also have to ask: “Where did this photograph come from?”

If video once reduced the distance between observation and reader, the video itself now needs a history of origin. Who recorded it? When? Where? On what object? Is there a sequence of original frames? Are there RAW files? Is there a specimen, a slide, an observation log, a repeated registration, or independent confirmation?

The evidence is no longer the image by itself. The evidence is the traceable path from the object to the image.

Material evidence has not become less important.

A microscope slide, fixed material, microbial culture, tissue sample, observation log, series of original images, or preserved object of study remains what it has always been: a way to return to the observation and examine it again.

But in a scientific article, the reader usually does not encounter the object itself. The reader encounters its representation: a photograph, a drawing, a table, a video frame, a description, or the author’s explanation. This is where the new question appears. If an image once helped bring the reader closer to the object, the image itself now needs a traceable origin.

The problem is not that material evidence has disappeared. The problem is that, in the public circulation of knowledge, it is increasingly not the object itself that circulates, but its persuasive digital shell.

Therefore, evidence is no longer the image itself, but the ability to trace the path from the object to the image.

Science spent a long time learning how to preserve observations.
Now it also has to learn how to preserve the provenance of observations.

FAQ

Does this mean photographs and videos are no longer useful?

No. They are still essential. But they should be treated as documented data, not as self-sufficient proof. Their value increases when the circumstances of recording, original files, sequence, specimen, or observation log can be checked.

Is this only a problem for AI-generated images?

No. AI makes the problem sharper, but the principle is older. Weak documentation, selective images, missing context, and overconfident interpretation existed long before generative AI.

What should be preserved together with an image?

As much context as possible: date, place, object, method, equipment, original files, repeated frames, related samples, microscopy slides, records, and any information that allows another person to reconstruct the path from observation to conclusion.

Can a conclusion be correct even if the documentation is weak?

Yes. A weakly documented conclusion can still be true. The problem is different: the reader has fewer ways to verify it independently and must rely more heavily on the author’s authority.

What changes for scientific communication?

Good communication is no longer just clear writing and strong visuals. It also requires provenance: showing where the data came from and how the conclusion was built from them.