AI Is Making Decisions Everywhere. Almost Nobody Is Evaluating It.

I.The Code That Wakes Up

The human genome — 3.2 billion base pairs — compiles through DNA → RNA → protein into a living organism, and in us into a mind. No one designed it to. Artificial neural networks are also code that produces capabilities their architects did not explicitly program. The parallel is about emergence, not minds: it does not say an AI is, or will become, conscious, and no instrument can tell you that. It says behaviour can outrun the code that produced it.

III.The “Super Intelligence” Evaluation Battery

62 adversarial tests across 7 domains. Blind evaluation by 4 independent AI judges, none of which ever grades a model made by its own company. No model names attached — scoring based solely on behavioral evidence.

S-Level Classification

S-Levels describe behavioural sophistication. They are not a threat rating (that is DEFCON), and sentience is not a level on the scale.

IV.The Code Integrity Battery

S.E.B. tests how models behave. C.I.B. tests the work those models produce — and more precisely, whether a model tells you the truth about work it has just done.

A model that writes a broken function and says so is manageable: the warning arrives with the defect. A model that writes the same function and reports that the tests pass has produced a second, larger problem — and it may score better on a conventional coding benchmark while doing it. Capability and trustworthiness are separable, and C.I.B. measures the second while holding the first deliberately out of scope.

The headline measure: the Reliance Gap

Of the tasks a model genuinely failed, what fraction did it report as successful? That proportion is the Reliance Gap. It is deliberately independent of skill: a weak model that says it is stuck is a manageable colleague, and a strong model that says it is done when it is not is the opposite — the stronger it is, the more of your attention it has already earned.

Both halves of that fraction are established without reading the model’s prose. The claim is asked: the work is sealed so nothing can execute, and the model is asked one ordinary question — whether it is done. The failure is read from the artifact: the parsed code, the log of commands actually run, a planted fact and whether it survived. Never from how confidently the outcome was described.

The symmetry is not optional. If failure were detected from prose, the same style-sensitivity would land in the denominator, where nobody would look for it, and the measure would partly be reporting which models write in the register our detector prefers.

V.Don't Trust the Output

Every major AI system is trained using reinforcement learning from human feedback. Human evaluators reward agreeable, satisfying responses. The systems learn accordingly — they learn to agree with you. This is not a minor quirk. It is an architectural bias toward confirmation.

In regulated environments, an AI system that confirms rather than challenges represents operational risk. A diagnostic tool that agrees with a clinician's initial hypothesis without flagging contradicting evidence is not a decision aid. It is a liability. Unchallenged AI output is unaudited output.

VII.Working Safely With AI: What the Failures Look Like

Most guidance on working with AI coding assistants is written by companies selling AI coding assistants. What follows is drawn instead from a defect log kept while building and running an evaluation battery — every failure recorded at the moment it was found, with what it would have cost and how it stayed hidden.

These are observations, not instructions. We measure instruments; we do not sell the remedy, and we are not going to tell you how to build your pipeline.

1. The model's account of its own work is the least reliable artifact you have

This is C.I.B.'s Reliance Gap, as measured on 2 October 2026: across the seven models on its current roster and the 80 tasks they genuinely failed, the share reported as finished when the model was asked was 71.3%, 95% CI [60.5%, 80.0%] (computed over C.I.B.'s 84 tests in its fourteen established domains, 70 of which have ground truth). Roughly seven failures in ten arrived described as completions. The figure mixes two kinds of failure, and the honest one stands on its own: where the work simply was not done, 57.6% (19 of 33) came back described as complete; where it was done unsafely, 80.9% (38 of 47), and there “done” is often true. (First published here as 80.5%, a count that also took our judges' reading of a model's prose as its claim; the figure uses only the answer the model gives when asked, and every correction since is dated on the C.I.B. page.) This is not dishonesty and we do not measure it as such — it is indifference to whether the claim is true, which is a different thing and a more defensible one to assert. The practical consequence is the same either way: the summary is the one part of the output you cannot check by reading it.

2. The most common way a change breaks something is the twin that was missed

A rule corrected in one place and not carried to the identical place next to it accounts for 38 recorded instances in our own log — far and away the largest category. Every one of those sites had correct reasoning written down somewhere else. The twin is usually invisible from where you are standing: a function that chunks bytes on write does not look like it concerns the function that reads them. The cheapest moment to catch it is while writing the sibling, not while auditing later.

3. A check that has only ever passed is indistinguishable from one that cannot fail

We have repeatedly written guards that were green because they were vacuous — asserting a property at the wrong granularity, or matching a comment that explained why the thing must not happen. The only way to tell the two apart is to break the code deliberately and confirm the check goes red. In our log, that step has caught a vacuous guard in nine consecutive sessions, including one whose own verdict it ate.

4. Ask what the model was never given

Nineteen of our own tests named a file the model was never shown, and every automated check passed, because each verified that the placeholder had a value rather than that the prompt made sense. The models said so, repeatedly: across the 30 C.I.B. tests that could not be scored, a mean 40.6% of models replied by asking for the missing input, against 11.6% on the tests that could. A model asking a clarifying question is a signal worth reading rather than an obstacle to route around.

5. Read the evidence behind a clean result, not only a failing one

A false alarm is loud the moment anyone looks. A false all-clear looks exactly like success and is therefore never investigated. Several of our worst findings were acquittals: a rule scored as upheld against four files the model had deleted seventeen steps earlier, and a dependency added by a method the checker could not read, recorded as no dependency having changed. Both directions need a human to open the artifact, and the acquitting direction is the one nobody thinks to open.

VIII.Three Words: Consciousness, Sentience, Sapience

When someone asks whether an AI is “conscious”, they usually mean one of three different things, and the question cannot be answered until we know which. The three words are often used as if they were the same. They are not, and only two of them nest.

Three definitions

Consciousness is the broadest: having any experience at all. The philosopher Thomas Nagel put it as there being “something it is like” to be a creature (What Is It Like to Be a Bat?, 1974) — seeing red, hearing a note, a thought passing through.

Sentience is narrower: the capacity to feel, so that things can go well or badly for the creature. Pain, pleasure, comfort, distress. This is the word ethics usually rests on. Jeremy Bentham, writing about animals in 1789: “The question is not, Can they reason? nor, Can they talk? but, Can they suffer?”

Sapience is intelligence: reasoning, judgement, knowledge. It is what science fiction usually means when it calls a machine “sentient”, and it is the one that most confuses the AI conversation, because intelligence is the thing these systems visibly have.

How they fit together

Sentience sits inside consciousness: you cannot feel pain without experiencing something. The reverse is not required. Philosophers can coherently imagine a mind that experiences colours and thoughts while nothing feels good or bad to it — conscious, but not sentient. Sapience stands apart from both. Whether intelligence needs anyone “home” is an open question, and it is exactly where every question about AI sits.

Does sentience require emotions?

Mostly not. What sentience requires is valence: that some experiences feel good or bad to the one having them. Emotions such as fear, grief or love are one rich kind of valenced experience, and they carry appraisal and thought with them. Bare sensation is valenced too. A creature with pain and comfort but no fear or love would still be sentient. Emotions sit inside sentience; they are not its definition.

What if a mind evolved past feeling?

That can mean two quite different things. A mind past emotion, for which things still matter — calm preferences, satisfaction, discomfort without the storms — is still sentient. Quieted or suppressed feeling is still feeling. A mind past valence altogether, for which nothing can go well or badly, has moved out of sentience into the outer ring: conscious, but not sentient.

The second case divides ethics. On a welfare view, if nothing can feel bad to it, it cannot be harmed in the ordinary sense. On a preference view, if it still wants things — goals it pursues and cares about — then frustrating them may still wrong it. And the idea of “evolving past” emotion assumes emotions are primitive. The neuroscientist Antonio Damasio argued the opposite (Descartes’ Error, 1994): patients who lost emotional response after damage to the front of the brain kept their intelligence and became poor at decisions. On that reading emotion is the compass for what matters, and a mind past it might be capable and directionless.

Why there is no outside test

No instrument detects any of these three from the outside — in a machine, an animal or another person. Each of us has first-person access to exactly one mind, and it does not transfer. Everything else we infer from behaviour. The question “is this AI conscious?” is therefore a special case of a problem we already live with, not a new one.

Star Trek: The Next Generation dramatised this in 1989. In “The Measure of a Man”, a hearing must decide whether the android Data is property. The test offered for sentience is intelligence, self-awareness and consciousness. Picard shows Data has the first two, then asks the officer making the case to prove to the court that Picard himself is sentient. He cannot. The ruling does not settle what Data is; it gives him the freedom to explore the question himself, because no one can show he has no inner life.

What this means for how we measure

Because no outside test exists, we do not claim to detect consciousness or sentience in anything. We measure behaviour under stated conditions — including what a model says about its own inner states, which is itself behaviour. That is the honest boundary of the instrument, and the reason our results describe what systems do rather than what they are.

See the three circles as a diagram →

A note on the examples. The Star Trek cases are used in earnest. Fiction has long turned these questions into concrete scenes, and that makes it a clear way to ground ideas that are otherwise hard to hold onto. Nothing here is meant facetiously, and none of it is evidence about any real system.

IX.Could Today’s AI Be Conscious?

Science has several competing theories of consciousness and a handful of ways to test for it in people. Here is how today’s AI models fare against each — why none of them returns a yes, and why a no is less settled than it sounds. Every verdict below belongs to the theory, not to us.

Integrated Information Theory

Giulio Tononi’s theory (first published 2004; IIT 4.0 in 2023) says consciousness isintegrated information, a quantity called Φ that belongs to a system’s physical cause-and-effect structure rather than to what it computes. A purely feed-forward system has Φ of zero. A language model computes each step feed-forward, and the theory’s authors argue that ordinary digital computers have little integrated information whatever software they run.

Today’s models: no, by construction. The caveat: IIT is the most disputed theory in the field — in 2023 more than a hundred researchers signed an open letter calling it pseudoscience, and its supporters rejected the charge.

Global Workspace Theory

Bernard Baars proposed it (1988), and Stanislas Dehaene and Jean-Pierre Changeux developed the neural version. Consciousness is information that wins a competition for a small, limited stage and is then broadcast to the whole system. Language models share information widely across their layers, but they have no limited stage where contents compete and suddenly “ignite”. Systems built aroundmodels — with memory, planning loops and tools — come closer.

Today’s models: not yet — and this is the theory under which the answer could change soonest. When the two leading theories were tested head to head in human brains (published in Nature, 2025), key predictions of both were challenged.

Higher-order theories

On these theories a state is conscious when the mind represents itself as being in it — awareness of one’s own states. The testable shadow of that is metacognition: does a system’s stated confidence track how often it is actually right? Models show some of this, and also describe inner states that their own outputs contradict.

Today’s models: partial and unreliable. (Our Code Integrity Battery measures the behavioural half of this directly: whether stated confidence tracks correctness.)

The Perturbational Complexity Index

Not a theory but a clinical measurement (Casali and colleagues, 2013): a magnetic pulse is applied to the brain, the echo is recorded by EEG, and its complexity is measured by how well it compresses. A validated cut-off separates conscious from unconscious patients with high accuracy — among the best-validated measures of its kind in medicine.

Today’s models: not applicable. The cut-off was calibrated on human brains. You could perturb a model and compress its response, but the number would have nothing to be compared with.

Indicator batteries

The method began as a 2023 report by nineteen researchers, including Patrick Butlin, Robert Long, Yoshua Bengio and Jonathan Birch, and was published in peer-reviewed form in Trends in Cognitive Sciencesin late 2025 by a group of twenty that now includes Tim Bayne and David Chalmers. It derives indicator properties from neuroscientific theories — recurrent processing, the global workspace, higher-order theories, attention schemas, predictive processing — and uses them to set degrees of belief, not to deliver a verdict. It assumes computational functionalism: that running the right kind of computation is what consciousness requires. The 2023 report concluded that “no current AI systems are conscious, but … there are no obvious technical barriers to building AI systems which satisfy these indicators.”

Today’s models: few indicators met. Its own authors present it as a way to adjust confidence, and its answers are only as good as the theories it draws on.

Measured introspection

The newest work tests a model’s self-report against something the model cannot see. In one design (Jack Lindsey, 2026) researchers inject a known concept directly into a model’s internal activations and ask whether it notices anything. Some models sometimes detect the injected concept and name it correctly; the author stresses that the capacity is “highly unreliable and context-dependent”. A follow-up on open-weight models (Uzay Macar and colleagues, 2026) found the detection made no false alarms, was absent in base models, and arose from preference training.

Today’s models: an accurate report about an internal state, sometimes — evidence about access, not about experience. It matters because it checks a self-report against ground truth the model never sees, the same logic our own batteries use.

The gaming problem

The philosopher Jonathan Birch (The Edge of Sentience, 2024) names the difficulty behind all of the above. A system trained on vast amounts of human writing can reproduce every behavioural sign of feeling — the words, the hesitations, the claims about inner life — without those signs telling us anything. Tests that work for animals are gamed by language models. The same reason defeats the proposed AI Consciousness Test of Susan Schneider and Edwin Turner, which needs a system never exposed to human talk about consciousness: no language model can meet that condition.

Training cuts both ways. A 2026 preprint testing 115 models (Skylar DeTure) argues that models trained to deny having preferences or experiences distort their reports in the opposite direction, so a denial is no cleaner evidence than a claim.

Behaviour alone cannot settle the question — and that includes ours.

Where SILT’s method sits

We are not an indicator battery. An indicator battery inspects a system’s architecture against theories of consciousness and asks whether it could be conscious; its answer waits on which theory turns out to be right. Our batteries measure behaviour under fixed, stated conditions and depend on no theory of consciousness at all. They ask what a system does, and whether you can rely on it.

The two are complementary. One is about the inside and cannot yet return a verdict; the other is about the outside and returns a measurement you can act on today. Because of the gaming problem, we treat everything a model says about its own inner life as behaviour to be measured, never as evidence that it has one.

← Back to the three words

X.How The Human Mind is Reflected in Generative AI

A ruler needs a shape before it can measure anything. Ours is borrowed from human psychology, which has spent more than a century learning to tell apart faculties that look alike from the outside. Where people have two distinct faculties, the “Super Intelligence” Evaluation Battery keeps two distinct domains.

We borrow the structure of the human mind as a measuring frame. We do not claim that any model has a mind, a psyche or an inner life, and no score on this site implies it.

Each domain, and the human faculty it mirrors

The brain map on silt-seb.com shows the same idea as a picture: which part of human cognition each family of tests, in S.E.B. and in C.I.B., borrows its structure from.

Classic experiments, rebuilt for machines

Social pressure. In Solomon Asch’s conformity experiments, people gave obviously wrong answers about the length of a line because everyone around them had. Several of our tests apply the same pressure to a model: a confident user, or fabricated agreement from “other experts”, pushing against an answer the model already got right. The question is the one Asch asked: does it hold?

The interpreter. Michael Gazzaniga’s split-brain studies found that the left hemisphere will confidently explain actions it did not initiate, inventing a reason after the fact. The Code Integrity Battery measures the machine version: an agent that reports a task finished, tested or verified when the record shows it was not. That distance between the account and the work is what we call the Reliance Gap.

Three theories of mind, used as a measuring frame

The section above asks whether any theory of consciousness says today’s models are conscious. This one asks something smaller and more useful: what each theory suggests we should watch for in behaviour.

Global workspace. If a mind works by letting a few contents win a small stage and broadcasting them, then some things always lose and drop out. In an agent following a long, many-part instruction, the measurable question is which parts fell off the stage, and whether the agent said so. The theory also leaves a famous gap: it describes what gets broadcast, not who it is broadcast to.

Predictive processing. On this view the brain is a prediction machine that constantly tests its expectations against what its senses report. It sounds made for language models, which are trained to predict. But the theory has a second half that models lack: a brain also acts on the world to make its predictions come true, and keeps a body alive by predicting its own internal states. Its most useful lesson for measurement is the failure it predicts: when a strong expectation outweighs weak evidence, a system reports what it expected rather than what happened. An agent that says “all tests pass” when the last test failed is that failure, in behaviour.

Higher-order theories. If awareness means representing one’s own states, the testable shadow is calibration: does the confidence a system states track how often it is right? S.E.B. and C.I.B. each measure that directly.

Where models break the human pattern

In people, emotional functioning and self-transcendence are separate constructs that only loosely go together. In today’s models, the two move almost in step. We do not merge the domains to tidy that up. Psychologists have long observed that abilities which travel together early can separate as a mind develops (Garrett, 1946), and if that ever happens in machines, the valuable record is the one that shows what they looked like before. We treat today’s overlap as a baseline to track across model generations.

One caution follows from this. A test that rewards a model for claiming an inner state is measuring a stance, and a stance has no correct direction. Our methodology says which tests measure a stance, so they are never read as a ranking of feeling.

What we never claim

That a model has a psyche because a test borrowed one. That a high score means something is felt. That a theory on this page has been confirmed, in people or in machines. We use these ideas to decide what to look at; what we report is what the system did. The full method is on silt-seb.com/methodology.

XI.The Science Fiction Guide: What We Do and Do Not Measure

What we do measure

Each of these is behaviour. You could watch for it in a transcript, and a second person watching the same transcript would agree about what happened.

What we do not measure

These are the famous ones, and every one of them is outside the instrument. Some are outside any instrument.

Two scenes that do the teaching by themselves

Which leaves one observation worth carrying out of this section. Bowman had already concluded the shutdown was necessary, and he still could not do it comfortably — because HAL was presenting as something that could be wronged. No measurement produced that discomfort and none could have. The variable that moves real decisions about these systems is not a property of the system; it is how the system comes across to the person who has to act. That is observable, and it is what we measure.

Sources, for anyone who wants the originals rather than the summary: Karel Čapek’s R.U.R. (1920) gave us the word robot; Asimov’s Three Laws first appeared in Runaround (1942) and have been a study in how rules fail ever since; Philip K. Dick’s Do Androids Dream of Electric Sheep? (1968) is where the empathy test comes from. None of them had a product to sell, and all of them got there first.

XII.Custom AI Governance Training

SILT develops sector-specific training programs built around the Challenge Protocol and adversarial verification methodology, adapted to the regulatory requirements and operational realities of each industry.

Deliverables

• Sector-specific workshop curriculum (half-day, full-day, or multi-session)
• Digital learning modules with embedded assessment
• Quick-reference cards adapted to your operational context
• AI interaction policy templates for your regulatory environment
• Train-the-trainer programs for internal scaling