On 15 October 2026, OpenAI policy counsel Shira Anderson reportedly told the US Judicial Conference’s Advisory Committee on Evidence Rules:
“AI images are only getting better, and I do not think that there will be a scientific way going forward to detect them.”
The remarks were reported by Reuters in a discussion about synthetic evidence. The quotation is also reproduced in public coverage of the meeting.
The broader discussion concerns an entirely legitimate problem. Generative models are improving. Human visual inspection is increasingly unreliable. Watermarks can become undetectable. Metadata can be removed. Different technology companies use different provenance systems. No universal mechanism has been demonstrated that reliably identifies every synthetic image produced by every generator after every conceivable transformation.
On that much, I largely agree.
But that is not the same proposition as saying there will be no scientific way to detect AI-generated imagery.
That claim is bullshit.
Synthetic-media detection is not solved. RHEM Labs has not built a magical detector. No forensic technique can provide certainty in every conceivable case. The problem with the statement is what it implies about forensic science.
Anderson also reportedly said that she did not believe verification systems would ever be bulletproof. That proposition is much easier to defend. But “not bulletproof” and “not scientific” are very different things.
Science does not require certainty
Forensic science generally does not ask whether a method can classify every possible specimen correctly for all time. It asks narrower questions. What are the competing propositions? What observable characteristics would we expect under each? Can those characteristics be measured? How strongly do the observations support one proposition relative to another? Under what conditions does that inference remain valid, and where does the method stop working?
An inability to answer a question universally does not make it scientifically inaccessible. It establishes the boundary of the inference.
That distinction matters enormously for synthetic media. The current conversation repeatedly collapses several different forensic problems into one deceptively simple question: “Is this AI?”
Our current research separates them into three analytical layers: origin, zoning and transposition. Origin asks how an image came into existence. Zoning asks whether different regions have different generative histories. Transposition asks what subsequent transformations may have obscured, weakened or altered the evidence.
Those questions require different evidence and different methods. Treating all of them as “AI detection” is one reason this field keeps tying itself in knots.
An image is a physical or computational artefact
A conventional digital photograph does not simply appear. Light passes through an optical system and interacts with a physical sensor. Charge is generated and measured. Electronic noise is introduced. Signals are converted, processed, demosaiced, corrected, compressed and written into an image.
Those processes leave structure.
Synthetic-generation pipelines are different. Diffusion models produce imagery through computational generation rather than physical photodetection. Their output still has measurable structure; that structure has a different origin.
This is one premise behind our work on Sensor Pattern Origin Testing — SPOT.
SPOT does not attempt to recognise six fingers, strange text or impossible shadows. It examines measurable properties associated with image formation and asks a narrower forensic question: how strongly is the evidence consistent with physical camera capture versus synthetic generation?
My published experiment in the Australian Journal of Forensic Sciences compared camera imagery from the Dresden Image Database with DALL·E-generated imagery. Noise residues were extracted and characterised using variance, skewness and kurtosis before the resulting populations were modelled statistically.
Across 100 randomly sampled runs, the paper’s Results section reports mean ROC AUC of 0.998 and mean accuracy of 0.985. These are different measures: AUC describes discrimination across thresholds; accuracy describes the proportion correctly classified at the applied threshold. The publication record also explains the discrepancy between the abstract’s accuracy figure and the Results section.
More important than the classification result was the underlying observation: the examined synthetic and camera populations exhibited measurably different statistical behaviour. Their noise-residue measurements permitted evaluation under competing origin propositions.
That is a scientific measurement. It does not prove that every future synthetic image will always be detectable, or establish performance against every generator, camera, editing pipeline or distribution history. Nor does a measured difference, by itself, establish which part of the pipeline caused it. Confounding explanations and generalisation require further experimental testing.
Our SPOT research page distinguishes the published experiment from the evolving prototype. It identifies unfamiliar cameras and generators, screenshots, recompression, editing and acquisition histories as open evaluation questions.
Those limits define the scientific work still required.
Provenance and forensic examination answer different questions
OpenAI’s public verification tool examines supported provenance signals, including C2PA Content Credentials and SynthID watermarks. It reports whether those signals are detected. OpenAI correctly warns that finding no supported signal does not establish that content was not generated with its tools.
That is sensible engineering. Metadata may disappear, watermarks may degrade, and content may come from another generator.
A Content Credential carries signed claims about an artefact’s origin or history. A watermark embeds a deliberately introduced signal. An examination of image-formation characteristics asks what the artefact’s measurable properties reveal about the process that produced it.
All can provide useful evidence. Each has its own scope and failure conditions.
OpenAI’s published explanation of its provenance signals is more nuanced than the sweeping statement reported from the committee meeting. It describes combining metadata and watermarks because each can address limitations of the other.
We agree with using complementary evidence. The disagreement is with the leap from the absence of universally foolproof verification to the predicted absence of scientific detection.
Those propositions are not equivalent.
The future generator is not the scientific proposition
If a detector learns superficial artefacts produced by today’s generator, and tomorrow’s generator removes them, there is an arms race. That approach has an obvious generalisation problem.
Forensic science offers another framing. What generative process is consistent with the evidence we can observe?
A camera, scanner, printer, graphics renderer and diffusion model do not create images through identical processes. Sometimes the resulting differences will be measurable. Sometimes subsequent processing will destroy the discriminating evidence. Several generative histories may coexist inside one image. Sometimes there will simply not be enough information left to distinguish the propositions.
An inconclusive result is a legitimate forensic conclusion.
Mechanism-based examination does not make a method immune to counter-forensics or future generators. A generator could imitate a measured characteristic, and a camera could photograph a synthetic image on a screen. Support for camera capture would then be compatible with a synthetic scene. The proposition being tested must be precise enough to make that distinction visible.
Images are evidentiary artefacts. Conclusions should reflect their consistency with competing generative mechanisms, with uncertainty carried through the interpretation.
Courts do not need a magic deepfake detector
I would be uncomfortable with a court relying on a machine that produces a green tick beside “REAL” or a red cross beside “AI”.
Courts need evidence. What material was examined? What proposition was tested? What characteristics were observed? What reference population was used? How was the method validated? What alternative explanations exist? What are the known error rates and limitations? How strongly does the observation support one proposition relative to another?
Those are familiar forensic questions. Synthetic media makes answering them more urgent.
Our work uses likelihood ratios to express relative evidential support. A likelihood ratio compares how probable an observation is under one model with how probable it is under another. Depending on the measurements and the suitability of the models, evidence may provide strong support, modest support or effectively no discrimination between the competing hypotheses.
It is not a probability that an image is “real”. Its interpretation depends on the propositions, reference populations and validity of the models for the material examined. The SPOT technical evidence page explains that distinction and the experimental limits.
That is more useful scientifically than pretending every image contains enough evidence to produce certainty.
The difficult future is real
None of this means OpenAI’s underlying warning should be dismissed. Generators will change. Images will be resized, recompressed, screenshotted and deliberately manipulated. Adversaries will try to suppress forensic signals. Hybrid imagery will become increasingly common. Some evidence will become irrecoverable and some questions will remain unanswered.
Our research programme treats origin, hybridisation, localisation, counter-forensics and transformation as central problems. Different decisions also demand different evidentiary strength: a lead for further investigation and a conclusion offered in expert evidence require different justification.
The response to harder forensic problems is to determine experimentally where evidence survives, where it does not, and what conclusions remain defensible.
The present experiment cannot refute a prediction about every future generator. Equally, the absence of universal verification today does not establish that all future scientific examination will be futile. That stronger claim requires evidence of its own.
So, wanna bet?
There probably will never be a bulletproof synthetic-media detector. Forensic science does not need one.
What we need are measurable signals, falsifiable propositions, appropriate reference populations, known limitations, calibrated uncertainty and conclusions that say no more than the evidence can support.
We already know that some examined forms of synthetic generation produce measurable forensic characteristics. We have measured them. We have published results.
Now the harder scientific work begins: determining how far those observations generalise, how they survive transformation, how hybrid imagery should be analysed, and where the limits actually lie.
Perhaps one day synthetic and physically captured imagery will become indistinguishable under every scientifically measurable characteristic available to us. If that happens, the evidence should demonstrate it.
Until then, declaring that there will be “no scientific way” to detect synthetic imagery is a prediction about the future, rather than an experimentally established conclusion.
And we’re quite happy to test it.