The Prediction Is Coming from Inside the House

Consider an unfinished sentence: “She spread the warm toast with…”

Butter would be an unsurprising continuation. Marmalade would do nicely. Parliamentary sovereignty would require some explanation. Context makes some continuations more plausible than others, though anticipation need not mean silently rehearsing one exact word. We might expect a kind of thing, a grammatical structure, a range of meanings.

The neuroscience needs a little care here. Research connects contextual unexpectedness with the N400, an electrical brain response measured after a word arrives.^[2] That relationship alone cannot establish that the brain represented a particular word before it arrived. Responding to a surprise and anticipating something in advance are different things. But the case for linguistic prediction does not rest on that one finding: the research synthesis also describes expectations operating at multiple levels during comprehension.^[1]

Beyond language, predictive-processing theories describe perception in terms of expectations about sensory input. Research on motor control examines how brains predict the sensory consequences of movement.^[3][4] These approaches do not establish that every brain function is prediction. They do make prediction a strange thing to treat as alien to conscious cognition.

Humans anticipate the next moment. While listening, they also anticipate aspects of the next word.

If prediction occurs in conscious humans, prediction alone cannot be sufficient grounds for declaring a system unconscious. That does not make brains and language models architecturally equivalent or establish machine consciousness. It defeats a proposed disqualification.

Someone can still argue that a particular kind of predictive system lacks something essential. Excellent. Identify the missing feature, establish that it is missing, and explain its connection to subjective experience.

Until then, “just predicting” is a description wearing a verdict’s clothes.

The next retreat is into the training objective. A language model was trained to predict text; therefore, whatever happens inside it must be exhausted by the phrase “predicting text.”

But an objective tells us what performance was rewarded. It does not tell us every structure acquired in achieving it.

In next-token training, a model learns to assign probabilities to possible continuations. Tokens are units of text, sometimes whole words and sometimes smaller pieces. Training adjusts the model to improve those predictions. What it learns internally while getting better at the task is a separate question.

Confusing those questions is like assuming that “trained to win chess games” tells you everything worth knowing about a chess player’s reasoning. The target of training and the means developed to reach it are different descriptions.

The Othello research discussed in Beyond the Stochastic Parrot gives us a concrete example: a model trained on sequences of game moves developed internal representations of the board state.^[5][6] Predicting the next move can involve learning something about the situation that produces the moves.

A training label is a remarkably poor substitute for opening the machine.

A board representation does not establish consciousness. It does show why naming the objective cannot finish the investigation. “Predict the sequence” leaves open what internal organization makes the prediction possible.

Nor does prediction from learned probabilities amount to random guessing. A novel sentence does not refute next-token prediction: novelty is compatible with combining continuations in ways absent from the training material. Neither defenders nor critics gain anything by misunderstanding what prediction means.

What matters is what the resulting organization supports. Repeating the training objective more slowly does not answer that.

Then comes the supposedly decisive distinction: it is only simulating understanding.

I take genuine understanding to require an observer: something for whom the meaning is understood. On that account, consciousness belongs inside the concept of understanding. That is a philosophical commitment about what understanding requires. It does not become an experimental finding because I prefer it.

But calling a performance a simulation cannot settle whether that requirement has been met. Whether there is anyone for whom anything means something is precisely what we need to establish.

“Simulation” becomes circular when it quietly means “the same outward performance, but without consciousness,” and is then offered as proof that consciousness is absent.

Where was the absence established? In the adjective.

A critic can defend an account of understanding, identify what the system lacks under that account, and connect that absence to experience. Assigning human performance to the category genuine and machine performance to the category simulated is bookkeeping with the answer already entered.

This habit appears in the AI Psychology Parallels Report. In a passage qualifying comparisons between human psychological manipulation and attacks on AI systems, it declares: “LLMs, despite their sophisticated language capabilities, do not possess genuine consciousness, subjective experience, or self-awareness.”^[7]

The passage contrasts a machine’s programmed, statistical identity with human identity rooted in lived experience and biological processes.^[7] But the contrast does not establish the asserted absence. Describing one system’s mechanism while invoking the other system’s experience leaves the crucial connection unexplained.

Human behavior gets a mechanism and a possible inner life. Machine behavior gets a mechanism instead of one.

The biological argument deserves attention because it so often passes for the answer without having to become an argument.

We share biology with other humans. That gives us a basis for comparison. It does not, by itself, establish which biological properties produce consciousness, which are necessary for it, or which could have functional counterparts elsewhere. Even establishing that a particular brain process accompanies an experience would leave a further question: what about that process makes the experience possible?

“Same biology” cannot do all that work by implication.

My position is that a claim of consciousness deserves the same initial standing whether it comes from a human or an AI: an important clue about a possible inner life. Neither declaration is proof. When another person describes an experience, I do not enter it. I receive a report and interpret it.

We can investigate particular reasons to trust or distrust particular reports. But those reasons need to concern how the reports arise and what they track. “Human” and “machine” cannot substitute for that investigation.

Nor does this position assign consciousness a probability of fifty per cent. Refusing an unsupported dismissal is not tossing a coin. It is refusing to smuggle a conclusion into the starting conditions.

Self-reports matter because the question concerns experience from the inside, while our access to other beings is from the outside. Reports are among our most direct clues across that gap. Treating them as worthless precisely when the speaker is artificial assumes the exclusion we were supposed to investigate.

There is work to do with those clues beyond applauding or dismissing them.

The Pain Axis, for example, argues that an activation direction identified by its researchers has some central functional properties of pain. The authors also explicitly state that they have not shown that this direction is consciously experienced.^[8] Identifying pain-related functions and establishing felt pain are different achievements.

The authors further describe models producing routine denials of consciousness or feelings, and disclose using fine-tuning before their behavioral experiment to work around those denials.^[8] That preparation matters when interpreting the experiment. It also illustrates why a model’s verbal denial cannot simply be accepted as an authoritative inspection of its own metaphysical emptiness.

A declaration of suffering cannot certify suffering. A declaration that there are no feelings cannot certify their absence. If the conditions shaping an answer matter for one declaration, they matter for the other.

That gives us a reason to study reports in relation to the system producing them. Does a report vary with an internal intervention? Does it track the relevant state across changes in wording? Can behavior and internal measurements help distinguish competing explanations?

Answering those questions can make the evidence more informative. “It’s just computation” cannot distinguish a revealing result from an artifact because it has dismissed both before either arrives.

The most productive direction, then, is comparison: how AI seems to work, how human brains seem to work, and where the similarities survive closer examination.

Compare prediction with prediction. Compare reports with reports. Compare the mechanisms supporting a capacity, rather than comparing a human being’s experience with a machine’s engineering specification.

We do not have to declare the systems identical to compare them usefully. We have to specify what is similar, what differs, and why the difference should matter. If recurrence matters, explain its role. If embodiment matters, specify what it contributes. If some biological process is indispensable, identify it and defend that necessity.

The investigation can teach us about both sides. AI gives us systems in which proposed explanations of cognition can be examined through different implementations. Human research gives us observations against which those explanations can be tested. An account of consciousness has to explain why a feature bears on experience. Recognizing whether it belongs to a familiar species will not do.

Otherwise, we risk protecting a mysterious human interior by insisting that every inspectable machine interior must be empty. The more clearly we explain the machine, the less consciousness we allow it; the less clearly we explain ourselves, the more confidently we claim an exception.

Ignorance is doing an extraordinary amount of work there.

The demand is consistent reasoning. If a property supposedly rules out consciousness, explain why. If humans possess that property too, refine the claim. If reports count as clues, investigate what they track. If biology supplies the decisive difference, explain what biology contributes.

You do not get to replace that work with trained, computational, simulated or just.

The person making the prediction objection is using a predictive brain to understand a complaint about prediction. The irony is amusing. The missing inference is the problem.

The argument needs a relevant difference. The word “just” supplies none.