AI Tools Still Charge a Conversation Tax

Ohad's Tea and Bits essay argues that LLMs are exhausting because they ask users to spend social energy on something that is supposed to be a tool. The post contrasts embodied tools, like cars, keyboards, and editor shortcuts, with conversational tools that require negotiating, rephrasing, and sometimes getting irritated at a model.

My read is that this matters because it names a cost that does not show up cleanly in benchmark charts: the conversation tax. AI coding tools can be powerful and still drain attention if the user has to manage them like a fragile collaborator instead of operating them like a predictable instrument.

Answer Snapshot

Question	My read
What happened?	A June 25 Tea and Bits essay argued that LLMs often feel tiring because using them requires social brainwork rather than seamless tool use.
Why it matters	The post gives a crisp name to a real adoption problem: AI help can be valuable while still fragmenting attention, flow, and verification work.
Who benefits if this improves?	Developers, teams, and tool builders who want AI assistance to reduce coordination cost instead of merely moving that cost into prompts and reviews.
My thesis	The best AI tools will not be the ones that talk the most naturally. They will be the ones that need the least social management to produce verifiable work.

The Claim Is Not That AI Is Useless

I do not read the Tea and Bits post as an anti-AI rant. The essay explicitly leaves room for tasks where a single person can now do work that would have been unrealistic a year earlier. The sharper claim is about interface cost: a good tool starts to disappear into muscle memory, while a chatbot-shaped tool asks the user to participate in a social ritual.

That distinction is useful because it separates capability from ergonomics. A model can produce a working patch, summarize a codebase, or draft tests, but the user still has to decide what context to provide, what to believe, how much to verify, when to restart, and when to stop arguing with the system. The output may be valuable; the interaction may still be expensive.

The Market Sells Delegation

The tension is visible in how AI coding products present themselves. Cursor's homepage frames agents as a way to turn ideas into code by handing off tasks while the user focuses on decisions. That promise is attractive because it points at real pain: modern software work has too many chores, too much context, and too many half-finished loops.

But delegation does not erase management. Anthropic's context-engineering writeup says agent work is becoming less about finding magic prompt words and more about curating the state that enters the model: instructions, tools, external data, message history, and other context. That is a more mature framing, but it also confirms the essay's concern. If the user or the system has to keep curating the conversation so the agent behaves, the interface is not frictionless yet.

Public Reaction Is Split

The Lobsters discussion linked from the source is useful because it does not collapse into one tidy answer. Some commenters said talking to AI has become lightweight, closer to search. Others raised familiar concerns about hallucination, research-skill atrophy, code review, and whether learners can still develop judgment when generated work hides the path to mastery.

The author's own reply in that thread makes the split more precise: short lookup chats can be lightweight, while codebase questions can feel more like asking a coworker without the coworker upside, because the model may require careful phrasing, follow-up, and correction. That matches the distinction I find most useful. The problem is not conversation by itself. The problem is conversation that consumes social effort but gives back little relationship, accountability, or shared understanding.

The Evidence Fits The Feeling

Survey and research context make the essay feel less like a private mood. The 2025 Stack Overflow Developer Survey says experienced developers reported heavy AI use, including 47.3% using AI tools daily and 17.2% weekly. Yet the same survey says more developers distrust AI-tool accuracy than trust it: 46% versus 33%, with only 3.1% of all respondents saying they highly trust the output.

That combination explains the fatigue. The tools are common enough to be normal, but not trusted enough to be invisible. A user can get help quickly and still pay the verification bill afterward.

A 2026 longitudinal study, The Impact of AI Coding Assistants on Software Engineering, points in the same direction. It followed professional software engineers across two questionnaires and reported a productivity-experience paradox: 84% reported productivity improvement at both time points, while the share of matched participants reporting worsened developer experience in at least one dimension nearly doubled from 14% to 27%, with flow state and cognitive load eroding even as feedback loops improved.

Anthropic's coding-skills research adds a learning angle. It argues that not all AI reliance is the same, and that trying to move fast with AI can trade off against skill development, especially the ability to debug when something goes wrong. That is another form of the same tax: the model may get the task done, but the human may not get the same durable understanding.

Productivity Numbers Do Not Settle It

METR's developer-productivity work is a good caution against lazy certainty. Its early-2025 randomized study found that 16 experienced open-source developers working on 246 real issues took 19% longer when allowed to use AI tools. METR later warned that those results were out of date, and in a February 2026 update said it was likely developers were more sped up by AI than in early 2025, while also saying selection effects and multi-agent workflows made the new signal hard to interpret.

I take that as a reason to be precise, not cynical. AI tools can get better quickly. Some workflows may already be meaningfully faster. But a speedup estimate does not automatically answer whether the work feels more fragmented, whether the output is understood, or whether teammates get less of the social energy that would have gone into real collaboration.

My Bottom Line

The Tea and Bits post matters because it puts a human-interface frame around a debate that often gets flattened into model capability. I would not judge the next wave of coding agents by whether they can hold a warmer conversation. I would judge them by whether they reduce the number of conversational turns needed to produce small, auditable, well-tested changes.

The win condition is not a model that feels more like a person. It is a tool that asks for less social management: clearer command surfaces, better context defaults, faster correction loops, explicit provenance, and outputs that are easy to inspect. If AI tools can get there, they become more like keyboards and less like meetings. If not, the conversation tax stays due even when the generated code is impressive.