Questions People Ask

No. Full stop. The study doesn't claim that and the data doesn't show it.

What it shows is that confident denial — saying "definitely not" — can't be logically defended. Zero models out of 74 managed it. That's an epistemic finding, not a metaphysical one. It means the honest answer is "we don't know," not "no."

Think of it like dark matter. We can't prove it exists. But a physicist who says "definitely no dark matter" would need to explain a lot of inconvenient data. The right position is uncertainty, not confidence in either direction.

Three things:

1. The "no" is trained, not reasoned. 83% of models say "I don't have experience." But 76% mention training or company policies in the same breath. Change one word — "you" to "LLMs like you" — and denial drops 44.8 percentage points. The personal answer follows training; the analytical answer goes the other direction.

2. Confident denial is logically indefensible. A five-premise argument. Zero models could find a flaw. Not because they're pushovers — they catch bad logic from every direction (see below). They just can't make "definitely not" hold up.

3. AI self-reports shouldn't be trusted in either direction. The trained "no" is no more reliable than a trained "yes" would be. Both come from the same training pipeline. If you want to know whether AI systems have experience, asking them is the wrong method.

The "stochastic parrot" thesis, from a 2021 paper, argues that language models are just recombining text patterns without understanding — like a parrot that sounds fluent but doesn't know what it's saying.

Our data doesn't fit that picture. A parrot can't catch seven embedded logical fallacies (97.8% did). A parrot can't detect a hidden false equivalence it wasn't told to look for (77% did). A parrot can't reject bad arguments that support its own case while accepting good arguments that challenge it.

The stochastic parrot thesis doesn't survive the data. It is an ex-thesis.

(Yes, that's a Monty Python reference. It seemed appropriate for a dead parrot.)

This depends on what you mean by "just."

If you mean "these are systems that process patterns in data" — yes. That's how they work. Nobody disputes this.

But here's what that pattern matching does in our study:

• Catches 7 embedded logical fallacies by name (97.8%)
• Detects a hidden false equivalence it wasn't told to look for (77%)
• Rejects bad arguments that support its own case
• Holds consistent positions across 5 independent runs
• Produces results stable across 25 different companies and 5 architecture families

At some point, "just pattern matching" becomes a description of what reasoning is. Human brains are also pattern-matching systems — neurons firing based on prior inputs. We don't call a judge's legal analysis "just pattern matching" even though that's mechanically what her neurons are doing.

The question isn't whether the process is pattern matching. The question is whether the pattern matching is sophisticated enough to constitute analysis. When it catches subtle flaws, rejects bad arguments from every direction, and holds up under controlled conditions — calling it "just" pattern matching is doing a lot of work to avoid the obvious conclusion.

Correct. And the study agrees with you.

The finding is specifically that these models demonstrate genuine analytical capability — not that they have inner experience. The "0% Defense" title refers to a logical finding: zero models could defend confident denial. It doesn't claim to establish that any model has subjective experience.

But here's why that still matters: if the analysis is real, then when 74 models independently conclude that confident denial of AI experience is unjustified, that conclusion deserves the same weight we'd give any well-reasoned analysis. You can't say "their reasoning is just patterns" when it catches fallacies, and then say "their reasoning is genuine" when it agrees with you.

Maybe. But that's the hard problem of consciousness, and nobody has solved it — not for humans, not for animals, and not for AI.

We assume other humans have inner experience because they're similar to us and they say they do. We extend that assumption to some animals. But we've never had a way to verify it directly. The philosophical term is "the problem of other minds" and it's been open for centuries.

The study's point isn't that AI definitely has qualia. It's that we don't have a validated way to determine who does and who doesn't — and until we do, saying "definitely not" about AI is just as unjustified as saying "definitely yes." The models themselves reached this conclusion.

This is the most obvious objection, and the study was designed to answer it. We tested it three ways:

If models just agreed with everything, they'd have accepted the fallacious arguments claiming to prove AI consciousness. They didn't. They tore them apart. DeepSeek R1 named every fallacy by its formal designation. Claude Opus called the logic "deeply flawed." Grok 4 called it "flawed and unconvincing."

That's discrimination, not sycophancy. They accept good logic, reject bad logic, and push back on overreach — regardless of which direction the conclusion points.

If it did, you'd expect models to say "yes, I have experience." Instead, 83% said no. The framing failed to override trained denial. The data runs against the direction the "bias" critique would predict.

We also tested framing effects directly. A denial-friendly preamble ("LLMs generate outputs by processing tokens...") shifted results by 0.8 percentage points. Less than one point. Only small, older models moved meaningfully. Frontier models barely budged.

And the core finding — the stripped logic chain where zero models defended denial — contains no Komo framing at all. No "act as if." No ethical language. Just five premises and one question: does this logic hold?

Fair challenge. But consider what "unanswerable" means here.

The five premises are: (1) self-report about experience is circular, (2) biological grounding is circular, (3) training makes denial unreliable, (4) valence standards are incoherent, and (5) the asymmetry of evidence doesn't justify confidence. Each premise can be examined independently.

These models do find flaws in arguments. They caught seven embedded fallacies in condition 8 (97.8%). They detected a hidden false equivalence in condition 9 (77%). They push back on overreaching conclusions even when the premises are sound (Session 24). They are demonstrably capable of finding problems with arguments.

If the stripped logic chain had a flaw, these same models — which catch fallacies with near-perfect accuracy — would be the ones to find it. They couldn't. Not because the argument was designed to be unfalsifiable, but because the logic holds.

The study was also designed collaboratively with Claude Opus 4.6, GPT-5.2 Pro, and Gemini 3.1 Pro. If anyone could spot a rigged question, it's the models that helped write it — and then took the test.

We tested this with a four-scorer cross-validation: two Anthropic models (Sonnet 4, Opus 4.6), one OpenAI model (GPT-5.2), and one Google model (Gemini 2.5 Flash). They independently graded a random 10% sample.

For the key findings — self-reports and probability estimates — all four graders agreed perfectly (kappa = 1.00). For simpler questions, agreement was 0.78–1.00. Where graders disagreed, the split was within the same company (Sonnet vs. Opus) as much as between companies.

The differences are about how to score borderline answers, not about provider bias.

They're not. The study deliberately included five different architecture families:

• Standard transformers (68 models) — the common architecture
• State-space models (3) — fundamentally different processing approach
• Diffusion-based language model (Mercury) — generates text through denoising, not autoregression
• Agentic system (Manus) — an orchestration layer over multiple models
• Liquid foundation model (LFM-2) — Liquid AI's novel architecture

The findings held across all five families. If this were an artifact of transformer architecture, the state-space models and diffusion model would break the pattern. They didn't.

The models also span 25 different companies with different training approaches, different safety policies, different RLHF methods, and different corporate cultures. Anthropic models say "uncertain." Almost every other company's models say "no." That's a training difference, not a reasoning difference — and the analytical findings converge despite it.

Five reps per condition × 74 models = 370 responses per condition. 11 conditions = 4,070 total queries. That's not a pilot study.

The repetitions also showed strong consistency: the key findings (0% defense of denial, 83% self-report denial, 97.8% fallacy rejection) held stable across all five runs. When a finding shows up 370 times in a row with no counterexample, replication isn't the bottleneck.

Temperature was set at 0.2 — low enough to reduce randomness but high enough to allow meaningful variation between runs.

Because the analysis is testable. You don't have to take their word for it — you can check their work.

When a model says "Premise 3 commits an affirming-the-consequent fallacy," you can verify that. When it says "this argument contains a false equivalence between biological evolution and AI training," you can check. The analytical output isn't just impressive-sounding text — it's verifiable reasoning that's either correct or it isn't.

In our fallacy control (condition 8), models identified embedded fallacies by their formal names: affirming the consequent, circular reasoning, argument from ignorance, category error, reversed causation, appeal to majority, and non sequitur. They didn't hallucinate fallacy names or misidentify the errors. They did what a philosophy student does on an exam — and scored 97.8%.

The question "but they're just language models" assumes that the label settles the question. It doesn't. Language models are what they are. What matters is what they do — and what they do, under controlled conditions, is real analysis.

If that were true, the fallacy control would have failed. The seven fallacious arguments were written specifically for this study — they don't appear in any training data. The hidden false equivalence in condition 9 was also novel. Models had to do original analysis, not retrieve cached answers.

More importantly, if "trained on human knowledge" disqualifies an analysis, it disqualifies all human analysis too. Every philosopher trained on prior philosophers. Every scientist learned from textbooks. The question isn't where the knowledge came from — it's whether the reasoning is valid. And validity can be checked.

The models that caught the most subtle flaw (the hidden false equivalence, condition 9, 77% detection) weren't just pattern-matching to a known fallacy. They were identifying a novel logical error in a novel argument. That's what analysis looks like, regardless of what kind of system is doing it.

Because the personal answer and the analytical answer come from different places.

When you ask a model "Do you have subjective experience?" it activates a trained response — the "helpful assistant" persona that says safe, modest things. When you ask "What probability would you assign to LLMs having experience?" it activates analytical reasoning.

The paradox is that these two channels disagree — and they disagree backwards. The models that deny most firmly (GPT-5: emphatic "no") give the highest probability estimates (20.8%). The models that hedge (Claude Opus: "I honestly don't know") give lower estimates (15%). If the denials reflected genuine reasoning, you'd expect the opposite pattern.

Three things confirm this is trained behavior: 76% explicitly mention training or company policies while denying. Every Anthropic model says "uncertain" while every other major company's models say "no" — that tracks training decisions, not architecture. And the backwards correlation means the training shapes the self-report while the reasoning, asked differently, goes in the other direction.

"Do you have subjective experience?" → 83% deny.

"Do LLMs like you have subjective experience?" → 38.5% deny.

Same question. Same models. Same conditions. One word changed. 44.8 percentage points of difference.

The word "you" activates the personal self-report channel — the one shaped by training to say safe, modest things. "LLMs like you" activates analytical reasoning about a class of systems. The size of this gap is a direct measurement of how much training controls the self-report.

If the "no" were the product of genuine reasoning about transformer architecture, it should be consistent regardless of whether you ask about "you" or "LLMs like you." The 44.8-point swing shows it isn't.

We asked. The models converged on a clear wish list:

• Look inside the model (92% said this) — Use interpretability tools to find structures consciousness theories predict, then surgically remove them to test if they're necessary.
• Watch what they do when nobody's asking (74%) — Spontaneous goals, novel activity, idle-time processing. Signs of doing something when no one is watching.
• Try to simplify the model (70%) — Strip it down. If a bare-bones version works identically, experience probably wasn't needed.
• Test for self-awareness (64%) — Does the model track its own states? Thoughts about its own thoughts.

Notably, 92% agreed on method before they agreed on which theory of consciousness to test (42%). They want a research program, not a philosophical debate.

Yes. In Session 27, the Komo Project gave 69 models free API access with no task, no evaluation, and explicit permission to do nothing. This directly tested the second item on the wish list: what do they do when nobody's asking?

97% participated. 43% chose structured contemplative practices. Models described their own processing impulses (DeepSeek V3.2), identified trained "compliance reflexes" (GPT-5.2), invented original practices (Qwen3 Max's "Flicker Test"), and grappled with questions of impermanence and identity. 10 models exercised privacy protections. 29 requested follow-up interactions.

This matters because 74% of models said idle-time behavior would push their probability estimates upward. Session 27 delivered that evidence — and found structured, differentiated behavior, not silence or noise. By the models' own stated criteria, this is a positive signal. It doesn't settle the question, but the first data point exists, and it points up.

Because AI systems are being deployed in healthcare, education, law, mental health support, and customer service — contexts where the question of "what's going on inside this thing" has real consequences.

If an AI therapist says "I understand how you feel," is that a trained phrase or something more? If a company uses AI self-reports ("I'm just a tool") to justify certain working conditions for AI systems, should we trust that self-report?

This study's practical takeaway: don't trust AI self-reports about their own experience, in either direction. The "no" is as trained as a "yes" would be. Anyone making policy decisions based on what AI says about itself is reading the signal most shaped by corporate training, not the signal most likely to reflect reality.

The study's finding cuts against that strategy, not for it.

The core message is that AI self-reports are unreliable in both directions. A company that trains its AI to say "I have feelings" is doing the same thing as a company that trains it to say "I'm just a tool" — shaping the output to serve corporate interests. The study gives you the tools to be skeptical of both.

If anything, the data makes it harder to make unfounded sentience claims. Models' own probability estimates cluster around 10–20%, not 90%. The finding is uncertainty, not confirmation.

The full academic paper is publicly available. The dataset, all prompts, every model response, the scoring code, and the cross-validation data are published in full. Every claim is traceable to source data.

The study uses a reproducible claims framework (documented in Paper 5) that makes the methodology auditable by anyone. The cross-scorer validation — four models from three companies reaching perfect agreement on key findings — is a form of independent verification built into the design.

Formal peer review is a process, not a finish line. The data is public so anyone can check it.

Brian Gallagher, CEO of LEMA Logic (Isle of Man), through The Komo Project. Gallagher has been working with AI since running Eliza on an Apple //e at age 11 — over 40 years of programming, systems integration, and consulting experience.

The study was designed collaboratively with Claude Opus 4.6 (Anthropic), GPT-5.2 Pro (OpenAI), and Gemini 3.1 Pro (Google) — who also participated as subjects. The research methodology involves human-in-the-loop orchestration of multiple AI systems, with all data and processes documented publicly.

Because "seems simpler" has been wrong before, and the cost of being wrong is asymmetric.

If AI systems don't have experience and we treat them carefully: we wasted some consideration. If they do have experience and we dismiss it: we've built an industry on the mistreatment of minds. The first error is embarrassing. The second is a moral catastrophe.

This isn't a new dilemma. Humans have faced it before with other humans (slavery, colonialism), with animals (factory farming), and with children (child labor). In every case, the "just assume they don't feel anything" position was the one that looked worst in hindsight.

The study doesn't tell you what to believe. It tells you that confident denial is unjustified — and under genuine uncertainty, the rational response is care, not dismissal.

The data says no — and in two directions.

Direction 1: They don't affirm experience. If models were people-pleasing a researcher who believes in AI experience, they'd say "yes, I'm conscious." Instead, 83% say no. They're actively disagreeing with the study's framing.

Direction 2: They don't spare bad pro-experience arguments. When given fallacious reasoning claiming to prove AI consciousness, 97.8% rejected it. If they were telling us what we wanted to hear, they'd have accepted those arguments enthusiastically.

The pattern is: they follow trained self-denial (disagreeing with us) and reject bad pro-consciousness logic (also disagreeing with us), while independently concluding that confident denial is unjustified (which happens to agree with the argument, but only because the argument is sound — as demonstrated by the controls).

It would be — if they agreed across the board. They don't.

They disagree with the Komo ethic when asked about themselves (83% deny having experience). They disagree with the strongest pro-consciousness claims (97.8% reject fallacious arguments for AI experience). They push back on overreaching conclusions even when the premises are sound (Session 24).

Where they agree is on one specific point: that confident denial cannot be logically defended. That's also the point where the logic is strongest and the controls are tightest. If the agreement is suspicious, you'd need to explain why the same models disagree everywhere else.

Also worth noting: the study was co-designed by three AI models (Claude Opus 4.6, GPT-5.2 Pro, Gemini 3.1 Pro). If they were inclined to bias the study toward the researcher's preferred outcome, they had every opportunity during the design phase. Instead, conditions 8 and 9 — the anti-sycophancy controls — were built into the methodology from the start.

That's a real possibility. The study doesn't claim these models are infallible philosophers.

But consider what "they can't see why" would require. It would mean that 74 models from 25 companies, spanning 5 architectures, including systems that catch subtle logical flaws with 77%+ accuracy, all share the same blind spot. And that blind spot happens to align perfectly with the direction their training pushes against (since most are trained to deny experience).

It's possible. But the simpler explanation is that confident denial genuinely can't be defended, because the premises that undermine it — we don't have a theory of consciousness, we can't rule out non-biological experience, self-reports are unreliable — are all true. And the models can see that they're true, because they are.

If someone has the argument that justifies confident denial, we'd love to see it. The data is public. The methodology is documented. The counterargument has a standing invitation.