Every source behind the argument

The article this site is built on. It shows how rules written to police what AI systems produce can end up filtering private thought before a word is ever published.

A German court considered whether a company is answerable for what its AI generated search summaries say. However the appeals fall, the underlying question is the one this site keeps returning to: when providers can be held to account for outputs, does the safe answer become the narrow one?

The document that turned an old right into a working test. Shaheed sets out the three things freedom of thought protects, keeping thoughts private, refusing manipulation, and never punishing thought alone, and states plainly that all three are absolute.

The practitioner brief that came before the book. Alegre maps the same three part test onto the surveillance advertising business model directly: what your data lets a system infer about your inner life, how attention and emotion get engineered, and how inferred traits can follow you into policing and insurance. Written just before generative AI arrived, so read it as the baseline the systems on this site have since moved past.

The book that put the forum internum back on the map. Alegre runs the right to think freely from heresy trials to targeted advertising, and argues the inner forum is under quieter attack now than at almost any point in its history.

The theoretical spine of the first direction. Teo argues that manipulation has to be defined by the process rather than by proven effect on a particular person, because a right that requires each individual to demonstrate their own distortion cannot function at the scale these systems operate on. That move is what allows flattery and quiet steering to count at all.

A comparative law argument from the religious freedom tradition. Courts have handled coercion and forced medication. They have never handled a chatbot that nudges gently and never announces itself. Pin places AI in that third, untested category, and argues that a tool woven into how you think carries the same protection as the thought itself.

The clearest sign that an international institution has adopted this exact vocabulary. The brief walks through generative AI, corporate influence, and neurotechnology under the forum internum heading, and calls for human rights impact assessments across the life of an AI system. It stops short of naming any single regulatory instrument, which is where this project starts.

The first attempt at a genuinely global survey of freedom of thought as its own adjudicable right, spanning eighteen jurisdictions. Its editors have called the right a sleeping beauty, present in law for decades and rarely given practical force. Nearly four hundred pages, organised by country. None of them is organised around a single regulatory instrument, which is exactly the gap this project is built to fill.

An early attempt to map every AI harm onto the full human rights framework, right by right. Freedom of thought appears here as one entry among many rather than as an organising idea, which is itself worth noticing. The digital rights movement had the vocabulary years before this project's central claim, that Europe's technical rules quietly decide the right, came into focus.

The European Parliament's own in-house evidence base treats the information sphere as a geostrategic battleground, authoritarian states and corporate platforms both manipulating opinion, generative AI as the accelerant. It is a security briefing, not a rights brief, and freedom of thought is barely named in it. That absence, from the institution actually voting on these rules, is a finding in its own right.

Where the term cognitive liberty actually began, a full generation before anyone applied it to a chatbot. Boire, a lawyer, built the case around an old United States Supreme Court line recognising a right to read or watch what you please in the privacy of your own home, and argued the same reasoning protects a right to control your own consciousness. The case was about psychoactive drugs. The reasoning about private thought travels further than its original subject.

Boire's co-founder at the Center for Cognitive Liberty and Ethics gives the term its working definition: a two sided liberty, the freedom to alter your own mind and the freedom from having it altered without consent. Written for nanotechnology and cognitive enhancement, not for language models, but the two sided structure is the one this project keeps returning to.

Carries cognitive liberty out of United States constitutional argument and into the vocabulary international human rights lawyers actually use, privacy and belief rather than due process. A decade before anyone wrote about corporate AI and cognitive liberty in the same sentence, the translation into rights language had already happened once.

The case for cognitive liberty when the technology reads the signal directly. Farahany asks what mental privacy can mean once a device can infer mood, attention, and intent from the body itself. This is the neurotechnology register the same term has also been argued in.

The first time the world's governments agreed a common ethical standard for technologies that act directly on the brain. Not binding, but a signal that the mind is now treated as territory the law has to defend. and the clearest sign the neurotechnology register of cognitive liberty has reached international standard setting.

The sharpest recent statement of a structural argument: AI agents owe you no duty of loyalty, and a market built this way makes a rights respecting assistant a hard sell. Radsch calls AI a new architecture of cognition, whose defaults decide whether the next generation experiences its own mind as sovereign territory or as a managed site. Written from an American vantage point, with little on what the European Union has already legislated against manipulation, which the brief on this site picks up. Read against Boire, Sententia, and Walsh above, this is the third register the same term has been argued in, not the first.

The other direction, argued well. An over-cautious AI can act as a shadow censor; a truth-seeking one could sharpen how people reason. The clearest short statement of the second threat this site describes.

The nearest prior art to the term thought moderation, and the debt is acknowledged here. Marks takes the vocabulary of platform governance and turns it inward, building toward a right to receive subconscious information. His legal order is American and his subject is technology acting on the brain directly, which is where his account and this one part company.

A review of a substantial Oxford University Press collection on AI and human rights, useful mainly for what it confirms. Freedom of thought still turns up as one chapter among dozens in the academic literature, rarely the organising question. Good for tracing where the wider field has been. Not a substitute for reading any single instrument closely.

The Board's first evaluation of large language models, and the freshest evidence for the filtering direction. Across ten leading models, refusals of political criticism ran more than twice as high for repressive governments as for permissive ones. The Board says plainly that it cannot yet explain why, which is part of what makes the finding land.

The measurement the Oversight Board finding was missing. Across 145 political questions, refusal rates ran to sixty percent for models built in China against close to zero for models built elsewhere, and the gap narrows sharply on harmless topics, which means training data and market preference cannot account for the whole of it. The detail that stays with you is a refusal to give travel advice about one section of the Great Wall.