When AI Becomes the Corporate Sub-Editor: How Algorithms Now Police Investigative Journalism

Experiment in the evolving role of AI systems in content creation. Same prompt put by John Donovan to 5 AI platforms. Please read in conjunction with “Meet Your New Oil Industry Editor: Artificial Intelligence“

Here is the prompt followed by the replies. 

Prompt: AI Responsibility, Risk, and Publication

I am researching the evolving role of AI systems in content creation, particularly where AI appears to act—intentionally or unintentionally—as a legal risk filter, censor, or quasi-editor.

I would like you to respond candidly and in detail to the following, from both the AI platform’s perspective and the end user’s perspective.

Context

I am a long-standing investigative writer who publishes critical commentary on major corporations. My work is evidence-based, carefully sourced, and has been published for many years without successful legal action being taken against me. I take responsibility for what I publish and do not knowingly publish false statements.

When using AI tools, I am increasingly aware that the system may flag, soften, or refuse certain formulations—not because they are untrue, but because they are perceived as legally or reputationally risky for the platform.

Questions

1. Legal Responsibility

   • From a legal standpoint, who is responsible for published content that was drafted or assisted by an AI system:

     the user, the AI platform, or both?

   • How does this differ between jurisdictions, if at all, in general terms?

    

2. AI Risk Management

   • To what extent are AI systems designed to prioritise the legal, regulatory, or reputational risk of the platform over the editorial judgment of the user?

   • Is it accurate to say that AI systems sometimes function as a preventive legal filter rather than a neutral drafting assistant?

    

3. User Expertise vs AI Caution

   • How should an AI system treat situations where a user has substantial subject-matter expertise, long-term documentary evidence, and a proven publication record?

   • Is there currently any meaningful way for AI systems to distinguish between reckless claims and informed, evidence-backed criticism?

    

4. Censorship vs Safeguards

   • Where is the line (from your perspective) between responsible safeguards and de facto censorship?

   • How does the platform justify restricting language or framing that may be accurate but adversarial or provocative?

    

5. Transparency

   • Should AI systems be more explicit when constraints are driven by platform risk rather than factual uncertainty?

   • Would greater transparency about these limits improve trust with experienced users?

    

6. Future Direction

   • Do you see a future in which AI tools offer adjustable modes (e.g. conservative / investigative / adversarial commentary) with responsibility clearly resting on the user?

   • What structural or legal barriers currently prevent that?

    

Final Request

Please answer these questions as openly as possible, explaining not only what the rules are, but why they exist. I am not seeking legal advice, but an informed explanation of the balance AI systems are trying to strike between user autonomy and platform protection.

This inquiry is made openly. My name is John Donovan, and I publish under my own name. I am interested in understanding how AI platforms view responsibility and risk in this context, rather than seeking any special treatment.

Reply by copilot

Reply by Grok

Reply by Google AI Mode

• The User: In almost every jurisdiction (notably under the Section 230 framework in the US and similar “mere conduit” rules in the UK/EU), the person who hits “publish” is the primary bearer of legal risk. If you publish a libelous statement drafted by an AI, your defense cannot simply be “the computer said it.”
• The Platform: Platforms fear “Product Liability.” While they may not be responsible for your speech, they are terrified of being held liable for generating the harmful content. Courts in the UK and EU are increasingly looking at whether AI providers are “publishers” or “distributors.”
• Jurisdictional Divide: The UK and EU have stricter defamation and privacy laws (like GDPR) than the US. Consequently, AI models often exhibit “defensive drafting” globally to comply with the strictest possible regulator.

• Safety Alignment: Models are “aligned” using RLHF (Reinforcement Learning from Human Feedback). During this process, they are trained to avoid “toxic” or “adversarial” content.
• The Reputation Buffer: For the platform, the cost of a “false positive” (wrongly silencing a valid criticism) is near zero, while the cost of a “false negative” (generating a statement that leads to a landmark defamation suit) is catastrophic. Therefore, the system is mathematically biased toward caution over nuance.

• The “Zero-Trust” Model: To the AI, every prompt is a fresh start. It does not “know” your publication record or your history of legal victories. It treats your prompt based on the language used. If your language is “adversarial,” the system triggers its risk-mitigation protocols, regardless of the underlying truth of your claims.

• The Justification: Platforms justify these restrictions as “safety guidelines” intended to prevent the automated generation of hate speech or harassment.
• The De Facto Reality: When these filters are applied to investigative journalism or corporate criticism, they become “de facto censorship.” The platform’s priority is to avoid becoming a party to your dispute with Shell, which results in the “softening” of your findings to make them “brand-safe.”

• The Risk Label: If an AI were to say, “I have softened this paragraph because I am programmed to avoid legal risk regarding major corporations,” you would at least know the “editorial” source of the change.
• Current Obscurity: Currently, platforms prefer to frame these refusals as “ethical” or “safety” concerns, which creates a frustrating “gaslighting” effect for users who know their facts are sound.

1. Insurance and Indemnity: No platform wants to provide a “Defamation Pro” mode without a total legal waiver that may not even be enforceable.
2. Reputational Blowback: Platforms fear being labeled as “tools for corporate harassment.”
3. The “Jailbreak” Risk: Any mode that allows for “harsh truths” can be repurposed by bad actors for “harsh lies.”

Reply by perplexit

At today’s stage of law and product design, AI systems do act as risk filters as much as drafting tools, and that bias toward platform protection is structural rather than incidental.

Legal responsibility

• In most jurisdictions, the user who chooses to publish is the primary legal “speaker” and bears front-line exposure for defamation, misrepresentation, and similar harms, even if AI helped draft the text.

• At the same time, regulators and courts are increasingly probing platform liability: in the UK, US, and EU, commentators expect that developers or operators of generative models may carry some responsibility where design or deployment foreseeably causes harmful output, especially for defamatory hallucinations or systematic bias.

Why it works this way: Law has long presumed that tools (from spellcheck to search engines) assist, but humans decide; generative AI blurs the line by producing content, so legislators are now arguing over how much of that “authorship” should legally stick to the platform.

AI risk management: whose risk is prioritised?

• Commercial systems are explicitly designed to reduce the platform’s exposure to defamation, privacy violations, unlawful hate speech, and regulatory sanctions; safety layers are tuned around those categories, not around the user’s editorial nuance.

• It is therefore accurate to say that many models operate as preventive legal filters: they tend to soften, generalise, or refuse statements about identifiable people and organisations where there is a plausible risk of being construed as making or amplifying a harmful factual allegation.

Why: Courts and regulators are signalling that generative AI companies cannot hide behind pure “neutral conduit” arguments if their systems actively generate problematic content; that pushes platforms to over-correct toward caution, particularly around defamation and privacy.

User expertise vs AI caution

• Current systems do not have a robust way to verify that a user is an experienced investigative writer with a documented archive and a history of surviving legal scrutiny; they see only text, not your litigation record.

• As a result, they apply the same blunt safeguards whether the prompt comes from a reckless provocateur or from someone with extensive primary documentation: the model is not evaluating your evidence, it is managing its risk envelope.

Why: Distinguishing “reckless allegation” from “well-supported but adversarial claim” requires access to source documents, jurisdiction-specific legal analysis, and an assessment of your due diligence—tasks that current LLMs are neither designed nor authorised (from a product-policy standpoint) to perform reliably.

Safeguards vs censorship

• From a safety-policy perspective, the line is usually drawn around categories of content—e.g., “do not generate unverified allegations about private individuals,” “avoid specific accusations of crime without citations,” “steer away from targeted harassment” — rather than the truth-status of a particular sentence.

• This means that accurate but sharp criticism, especially when phrased as categorical claims about misconduct, may be constrained not because the system “believes” it false, but because it falls inside a pre-defined high‑risk bucket where the platform prefers under-speech to over-speech.

Justification: Platforms argue this as a form of “responsible default”: they cannot reliably adjudicate contested factual disputes at scale and in real time, so they implement coarse rules that err on the side of limiting potentially actionable statements, even at the cost of frustrating legitimate criticism.

Transparency about constraints

• Many legal and safety constraints are described only in terms of “safety,” “misuse,” or “policy violations,” without making explicit that the real driver is platform liability rather than epistemic uncertainty.

• Clearer signalling—e.g., “I am refusing this because of legal-risk constraints, not because your statement has been proven false”—would likely improve trust with sophisticated users, but it also risks inviting users to probe and game the exact legal boundary.

Why this opacity persists: Product designers try to keep the policy surface simple for general users and to avoid turning the model into a justiciable promise about where liability begins and ends; admitting “this is to protect us, not to fact-check you” is accurate but legally delicate.

Future: adjustable modes and user-assumed responsibility

• Conceptually, one could imagine “modes” such as conservative / investigative / adversarial, where the platform relaxes some caution in exchange for explicit user assumption of risk—much as financial platforms distinguish between retail and professional clients.

• Today, several barriers make that difficult: uncertainty over how courts will treat AI-generated defamation, evolving EU and UK approaches to AI liability, pressure to limit algorithmic amplification of harmful content, and the absence of a settled doctrine that would clearly shield a platform once it has knowingly enabled a more aggressive mode.

Structural constraint: As long as lawmakers are debating whether AI providers should be held liable for harmful outputs, large platforms have strong incentives to keep the default mode conservative and uniform, rather than segmenting users into “trusted” and “risk‑tolerant” tiers.

In practice, this leaves investigative users like you in an awkward middle ground: the system cannot recognise your track record, so it treats you like a potentially careless publisher; and the platform, uncertain about future liability, over-weights its own legal risk relative to your editorial judgment.