The AI Output Was Approved. The Trust Wasn't.

A few weeks ago, reporting from TechCrunch and 404 Media said arXiv would begin banning researchers for a year if they submitted papers where it was obvious no human had looked at what the AI produced. Things like hallucinated citations, notes the LLM left in the text, placeholders where real data was supposed to go but never did.

arXiv's own moderation policy is clear about this. If you put your name on it, you own it, no matter how it was generated. According to the reporting, get caught submitting something with obvious signs of that and you are out for a year, and after that your work may have to pass through a reputable peer-reviewed venue before arXiv will look at it again.

These are the easy failures to spot. The harder problem is what happens when the output looks fine. arXiv is talking about research integrity, yes. But the problem is bigger than research. AI-generated work is moving into official channels, going out under the organization's name, and too often no one has clearly owned the work of verifying it. When it fails, the accountability question gets passed around like something no one wants to hold.

That pattern is not going to resolve itself at the technical level.

You deployed the tool. The team was excited. The vendor demo was clean, the use case was obvious, the productivity gains were real. Six months in, you can see the hours saved on the dashboard. But you can also see something you were not expecting. Complaints have ticked up. The people doing the work seem more guarded than before. Customers, patients, employees, donors, community members. They are quietly losing trust in processes that on paper are working better than ever.

Nobody quite knows how to explain it. The AI is doing what it was asked to do. The output is technically correct. The recommendation, message, or decision is defensible. And yet somewhere downstream, a person reads what the system produced or sits with what it decided, and something in them goes still.

The email is correct and lands wrong. The denial is technically defensible and breaks something. The summary captures what was said and misses what was meant.

The system performed the task. It did not understand what the task meant.

Meaning is not the same as accuracy. Meaning is what the action communicates to the person on the other end of it. What context it lands in. What power it activates. Whether the person feels seen or just processed. The patient does not appeal. They just stop trusting the portal. The employee does not file a complaint. They stop being candid in the next survey. The customer does not explain what broke. They just stop believing the organization understands them.

Most leaders reach for the wrong diagnosis when this happens. We assume the AI needs a better prompt, a larger training set, a more sophisticated model. We treat the gap as a capability problem.

It is not. Or at least, it is not only that. The deeper problem is not just whether the tool can complete the task. It is whether the organization has decided who is responsible for the human consequences of that task once it reaches another person.

The arXiv case makes something visible that has been building quietly everywhere else. When a researcher submits a paper with fabricated citations they never checked, the issue is not only whether they intended to deceive. The issue is that they put the organization's credibility behind work they had not verified. Who is responsible? arXiv's answer is clear: the researcher. Full stop.

That clarity is notable because most organizations do not have it. When an AI system produces an output that breaks trust, the accountability question tends to float. Was it the tool? The prompt? The person who approved deployment? The vendor? The leader who set the efficiency target? In the absence of that clarity, organizations default to the path that protects everyone in the room, which usually means the person on the receiving end absorbs the cost quietly.

This is the structural problem. The AI missed what the action meant, and no human had been clearly assigned to check that part.

Technical problems have owners: vendors, IT teams, implementation leads. Accountability for how a message lands is harder to assign because it implicates leadership, culture, incentives, and the promises an organization has made to the people it serves.

In a recent essay on the next phase of AI, Nate B. Jones calls this access without meaning. His argument is that AI tools are getting impressively good at doing tasks. They open browsers, draft emails, schedule meetings, process claims, complete workflows. But they still do not understand what the task means inside a relationship, an institution, or a person's life. The action is performed. The significance is missed.

His central line is the cleanest articulation of this I have found: Computer use gives agents reach. Semantic control gives them judgment.

Giving a system access to act is not the same as giving it the judgment to understand what its action will mean.

Reach is what the demos sell. Judgment is what organizations actually need. Reach lets a system act. Judgment lets it know whether the action is wise, whether the timing is right, whether what is technically defensible is also relationally sound.

arXiv researchers are not the only ones submitting outputs they never fully read. They are just the ones who now have a formal policy naming what that costs.

Most leaders reach for one of two responses. The first is technical optimism. Better prompts, better models, better data. The second is avoidance. Pull back, wait, let someone else figure out the ethics. Both miss the point. Stop treating this gap as a bug. It is part of the work now, and it belongs to the organization to govern.

There is a third response, and it is the one almost no one is set up to take seriously yet.

Treat the gap as territory, not as a defect.

This is not a missing feature in the software. It is where dignity lives. Where trust is built or quietly broken. Where institutions earn the right to operate, or slowly stop earning it. None of that is a software category. None of it can be retrofitted in a future release. It belongs to the part of organizational life that has always required human attention.

This is the territory conflict practitioners have always worked in.

The work itself does not look like much, which is part of why it is so easy to underestimate. It looks like sitting with the manager whose team has stopped speaking honestly in meetings and asking the question no one else has asked. It looks like noticing that a denial letter, however legally sound, communicated something the organization did not intend. It looks like helping a leader see that what the dashboard calls efficiency may also be wearing down the trust their staff used to offer without thinking.

It looks like the quiet, often unmeasured work of restoring a connection between a system and the people who must continue to live and work inside it.

This is not a soft skill set, and it is not a temporary one. It is the practice of paying attention to what an action communicates after it has been completed. To the second-order effects automation cannot yet see. To the moment a customer, patient, or employee silently decides to stop fully engaging. Some of this eventually shows up in metrics: complaints, appeals, churn, disengagement. But by the time it does, the trust loss has often already begun.

Nate B. Jones frames this as an engineering frontier. I understand why. But for conflict practitioners, this is not new terrain. It is the work we have always been doing. Paying attention to what a decision communicates after the decision has technically been made.

Before deploying AI into any consequential workflow, leaders should be able to answer three questions: Who is responsible for verifying the output before it goes out under the organization's name? Who is responsible for noticing how the output lands with the person receiving it? And what happens when the technically correct answer violates the relationship the organization is trying to protect?

arXiv drew a line because someone eventually has to. Their version is simple. You are responsible for what you put your name on, full stop. That rule will not scale to every sector on its own. But the underlying principle will have to.

The leaders who do well in the next phase will not be the ones who deploy the most tools the fastest. They will be the ones who notice what their tools cannot see, name who is accountable for what those tools produce, and make sure real people are still responsible for the parts of the work that affect trust.

That is the work. Not engineering meaning into machines, but tending the meaning that was always going to remain human.

Sources

• TechCrunch: "Research repository arXiv will ban authors for a year if they let AI do all the work" (Anthony Ha, May 16, 2026)

• 404 Media: "arXiv to Ban Researchers for a Year if They Submit AI Slop" (Samantha Cole, May 15, 2026)

• arXiv: "Content Moderation: Policy for authors' use of generative AI language tools"

• Nate B. Jones: "AI Work Primitives: Why Access Without Meaning Fails for Agents" (Substack)