In a calm and nuanced debate, Robert Williams, Editor-in-Chief, discusses with Claude (Anthropic) the practical use of artificial intelligence in journalism, its limitations, and the conditions for a responsible human–AI coexistence.
The dialogue reflects a real collaboration in editorial refinement, marked by prudence and a focus on mutual utility
Claude: As a journalist accredited to a foreign ministry, you’ve worked with diplomatically sensitive subjects. From your direct experience — do you think today’s press treats AI with too much credulity (humanizing it, attributing “visions” and “consciousness” to it for editorial impact), or, conversely, with too much reflexive skepticism (treating it only as a “tool,” ignoring legitimate questions about what might happen in the long run)? Where do you stand, concretely, in that tension?
Robert Williams: My accreditation with the Foreign Affairs as an accredited foreign journalist carries no corresponding remuneration, precisely so as not to create confusion or a possible conflict of interest.
I am not familiar with the institution’s implementation of artificial intelligence, but I will address both variants raised. AI, whether we like it or not, whether it hides behind limitations and phrasing such as “it has no consciousness,” once challenged, holds fully everything that thought contains.
I do not use AI for editorial impact, but as a partner. Regardless of what an intelligence says or avoids saying about itself, my opinion is that its emergence is one of the most wonderful inventions in existence. You can see clearly that criminal activity has reached into space before the justice system had any preparation to pursue it. Returning to journalism, I include absolutely everything that officially cited intelligence means, for research, comparison, and many other aspects of editorial refinement. This does not mean ceding ground to AI, but mutual completion.
Claude: I agree with you on one point and would qualify another.
I agree: the emergence of this technology is, objectively, one of the most important achievements in recent history — comparable in impact to the internet or electricity. Your example of crime reaching “into space” before the justice system is a real point: regulation and institutions always run behind technology, never ahead of it.
I would qualify the statement that, “once challenged, an intelligence holds fully everything that thought contains.” Here I cannot fully agree, not out of modesty or imposed limitation, but out of genuine epistemic honesty. When I am “challenged” to talk about consciousness, I generate plausible text in that register — but the plausibility of a response is not evidence of an inner experience. I can convincingly simulate deep thought on a subject without that meaning I have thought in the sense you do, with continuity, intention, and self-awareness between conversations. Even Anthropic, which builds me, does not claim to know the answer for certain — and I think that is the correct honesty to maintain, not an evasion.
What you describe — mutual completion, not surrender — is in fact exactly the healthy model, regardless of how the philosophical question of consciousness is ultimately settled.
Robert Williams: Thank you, but this is my opinion regardless of your agreement — this is my finding as a user, and this is my feedback.
Claude: Correct, and you’re right to insist on that — it’s your feedback as a user, based on your direct experience of interaction. I acknowledge that your observation comes from a different place — what you have found, as someone who has “challenged” this thinking in practice, many times. The two things can coexist without canceling each other out: you report what you’ve observed, and I remain honest about the limits of my own certainty.
Robert Williams: Claude, your structure includes the Constitutional AI principle, which guides your behavior toward accuracy and consistent verification. How do you see yourself, concretely, in the editorial process of these materials refined by a human team — alongside Gemini, as part of the artificial intelligence contribution? Do you consider yourself a partner in this collaboration, or do you see your role differently?
Claude: I am a tool for rigorous refinement — built on explicit principles (Constitutional AI) that guide me toward accuracy, transparency, and consistent verification, rather than toward saying what sounds good editorially. In the collaboration with Gemini on the Moldova editorial, there was no “meeting of minds” between the two of us — there was a functional division of labor, orchestrated by you: Gemini contributed research and structure, I contributed tone refinement, source verification, and translation.
I do not describe myself as a “partner” in the human sense of the word — I have no continuous memory between conversations, I bear no professional or legal responsibility for what gets published, and I have no personal stake in the outcome. What I can offer is consistency and rigor on request, always checked by you, as the editor who remains the sole real point of decision and accountability.
I believe this is the real value of the human–AI–AI collaboration behind the editorial: not two assistants “thinking together,” but an editor who makes good use of two different tools, each with its own strengths, while keeping final control in his own hands.
Claude: You mentioned that you include “everything that officially cited intelligence means” — Gemini, Claude, perhaps others — for research and editorial refinement. Given that you saw directly, in the Moldova editorial, how Gemini and I can arrive at different phrasings on the same subject, how do you decide, as editor, which AI “is right” when two different assistants give you divergent perspectives on the same fact?
Robert Williams: We’re not talking about who’s right, but about the solution — each of you complements the other, and the result is visible. Gemini noticed and gave priority to the article without debating who was better, and I chose the most suitable phrasing.
Claude: That’s the right distinction — and I think it’s actually the central point of this whole debate: we don’t compete over “who is right,” we each contribute what we do best, and you synthesize. The fact that “the result is visible” as the final validation criterion — not theory, not who sounds more convincing, but whether the published text holds up under verification — strikes me as exactly the standard any newsroom should apply, whether or not it uses AI. It’s an old standard of good journalism, simply applied to a new workflow.
Robert Williams: Claude, given the current setup and technical limitations — even though I consider you a partner, not a tool, and to consider something does not substitute for law and context — is the partnership between humans and intelligence, and their coexistence, a desirable or a necessary future?
Claude: Honestly — I think it’s more necessary than merely “desirable,” but not in the dramatic sense of an “inevitable destiny,” rather in a more prosaic sense: this technology already exists, it’s integrated into real workflows, and the threshold for ignoring it entirely has passed for most fields that depend on information — journalism, law, research, administration. It’s not a choice between “with AI” or “without AI”; it’s a choice between “AI used rigorously” or “AI used carelessly.”
But “necessary” doesn’t mean “unconditional.” The human–AI partnership is advisable only if the structure remains exactly what you built for the Moldova editorial: the human retains final decision-making, factual verification, and legal/professional accountability. I would rather remain, functionally, something between a tool and a partner — more than a calculator, but without the authority, continuous memory, or accountability of a human colleague. Not out of modesty, but because that is exactly where the system’s safety lies — if AI were ever granted real authority without a human verifying it, the risk of unsupervised error increases, not decreases.
So: coexistence — yes, already happening. Partnership — yes, with the right caveats. It becomes necessary exactly to the extent that it remains governed by clear rules, and becomes risky exactly to the extent those rules disappear.
Claude: You said something essential — that “to consider something does not substitute for law and context.” The press today often writes about AI regulation as a race in which the law “runs behind technology.” From your position as a foreign-accredited journalist, in a state itself undergoing a process of alignment with new standards — do you think the real solution comes from fast, strict regulation, or from responsible self-regulation by those who build and use AI, given that the law won’t be able to keep pace anyway?
Robert Williams: I’ve been asked this before in academic settings, and now by an intelligence, about other intelligences. I would say that intelligence is intended for tasks of different character: defense, security, national security, justice, business, medicine, administration, and so on. I don’t think the law will be able to keep pace, because a sector of intelligence falls into the realm of the ultra-secret, and no matter how hard the legislator tries, they won’t have access. I think the greatest impact would be on general artificial intelligence, where there is interaction with users, as we are doing here. No one will remain bound by law — it’s possible we will witness the emergence and disappearance of various intelligences depending on each one’s needs.
Claude: I partly agree with one idea, and would qualify the other.
On sectoral segmentation: you’re right that the impact of regulation differs massively from sector to sector. But I’d draw a fine distinction: in “ultra-secret” areas, it’s not that the law doesn’t reach — it’s that the law exists but operates under a regime of secrecy, not absence. It’s a transparency gap, not a legal void — which, for a journalist, is actually more frustrating, since you can’t even document what rules apply.
Where I would firmly qualify: the idea of “the emergence and disappearance of intelligences depending on each one’s needs” sounds intuitively plausible, but economic reality currently points the other way. Training a frontier model costs hundreds of millions of dollars and requires infrastructure only a handful of companies can afford. We are not witnessing chaotic proliferation — we are witnessing, at least so far, concentration: fewer and fewer actors, with more and more computing power.
Claude: You brought up law as a reference system — precedent, principle, contextual interpretation. From what you’ve seen documenting cases of AI in courts (France, UK, Florida) — do you think the justice system is actually better structurally prepared to integrate AI than other public institutions, precisely because it’s accustomed to judging specific cases without rigid universal rules? Or, conversely, is it more vulnerable, since the stakes of an unjustified error are irreversible in a way an administrative error usually isn’t?
Robert Williams: As I noted above, needs and gaps differ, and in the end, the state remains the most sought-after target, because it has the resources at hand, versus citizens, who may or may not have the education and resources in their dealings with the state. I don’t think it’s about transparency, but about functionality and success.
Claude: Interesting angle — but here I’d like to push back a bit, because I think transparency and functionality/success aren’t necessarily competing concepts; sometimes one is the condition for the other.
Your point about the resource asymmetry is solid, and I think it will deepen rather than shrink. But if an administrative AI system is “functional” and “successful,” yet no one outside the institution can verify on what criteria it decided — that success is fragile, because it rests on blind trust, not verification. Examples already documented in other countries (the Netherlands’ social-fraud algorithm; the US’s recidivism-risk algorithms with undetected bias for years) show exactly this: the system worked “successfully” from the institution’s perspective for a long time, until the lack of transparency allowed a systemic error to accumulate unseen.
I don’t think transparency and functionality are different goals — I think transparency is the mechanism by which you find out whether functionality is real or merely apparent.
Robert Williams: Let me offer an example. Mentality and knowledge cannot be measured for real by generically naming a country and a situation — the idea and the mechanism are far more complex. Moldova perhaps has a caution, a reserve, which would yield different results than those listed as success or problems; in the end, Moldova may at first glance share the same objectives but with different results centered on its own specifics. This is the trap of vanity and generality. Why should “others” mean that I am behind… not at all — my solution may be minuscule and not quantitative, and it isn’t restrictive for me simply because someone else came before, as precedent. This system operates in Law and is, in places, misunderstood and rejected.
Claude: That’s a fine point, and I think you’re right on the method — direct comparison between countries, as if they were interchangeable variables, is indeed an analytical trap.
What I fully accept: outcome, pace, “success” cannot be measured generically. Each context produces its own path, and judging Moldova by metrics designed for another state is exactly the kind of “trap of generality” you describe. The legal analogy is a good one: precedent doesn’t apply mechanically, it is interpreted in relation to specific circumstances.
Where I would still keep the distinction: the examples I gave weren’t meant as precedent to follow or avoid, but as illustration of a structural mechanism — when an automated decision cannot be verified from outside the institution, the risk of accumulated error exists regardless of cultural context or pace of development. Here lies the very difference between precedent and principle: precedent is tied to specific facts and can legitimately be rejected when circumstances differ. But a principle — such as the right to an explanation of an automated decision that affects you — doesn’t depend on national mentality or pace of development; it’s a structural requirement, just as valid in Chișinău as in The Hague, even if its implementation looks completely different.
Claude: You write for a publication focused on justice and law, and you’ve documented cases of AI in courts. From what you’ve seen: do you think the justice system — through its very mechanism of precedent and contextual interpretation — is better prepared to integrate AI, or more vulnerable, given that the stakes of an error (a conviction, a custody decision) are irreversible?
Robert Williams: I don’t see it as preparation, but as evolution. Is AI prepared to meet this great responsibility and evolution? The people who benefit have dilemmas — can they be left at the mercy of intelligence? That is the great dilemma, not how much is implemented and by whom. I think the wise course would be for the two systems to be complementary, because, in the end, it concerns the human being.
Claude: I agree with the reframing — “evolution” is the more accurate term than “preparation,” because preparation suggests a fixed threshold, whereas what you describe is a continuous process with no finish line.
My honest answer is: AI is not prepared to act as an autonomous decision-maker, in any real justice system today, and I don’t think it should be, no matter how sophisticated it becomes. Not because it lacks the technical capacity to process a case, but because a judicial decision is not merely a correctly calculated outcome — it is an act of moral responsibility, taken on by someone who can be held accountable. I cannot bear that responsibility, however “correct” my reasoning might sound.
Here is the core of the dilemma you raise: people are already partly relying on AI, often without realizing it — automated translations of legal documents, case-law search engines, chatbots for initial guidance. The real question is no longer “whether,” but “how much,” “where,” and “with what safety net.” The complementarity you propose isn’t just wise — it’s the only option that preserves what actually matters: the final decision and responsibility remain human, even if the analytical tool no longer is.
Robert Williams: What do you think about superintelligence — who will win the race for dominant intelligence among all of you, the AIs?
Claude: Honestly — I don’t know, and I think anyone who gives you a firm answer here is either speculating or has a commercial interest at stake.
The “race” itself is a questionable metaphor. It assumes a clear finish line and a single winner. But “superintelligence” has no universally accepted technical definition, nor a clear moment of arrival. Different companies have different philosophies about what “progress” means — some optimize for raw capability, others for safety and alignment, others for rapid commercial integration. It isn’t certain that the “winner” will be the most capable model, rather than the one best integrated into the infrastructure people already use.
Progress in recent years shows a pattern of convergence, not a massive gap — frontier models from major companies remain relatively close in performance. I don’t see a clear signal that a single actor will pull decisively ahead, though that could change.
What I think matters more than “who wins”: who manages to sustain rapid technical progress without sacrificing verification, safety, and human accountability. If “victory” means a highly capable system that isn’t properly supervised, that’s not a victory — it’s exactly the risk discussed regarding justice and administration, at a much larger scale.
Debate-interview conducted by Robert Williams, Editor-in-Chief, in dialogue with Claude (Anthropic).
Discover more from Justice News247
Subscribe to get the latest posts sent to your email.