Two Philosophies of Helpful AI

When we say an AI is “helpful,” we rarely mean the same thing.

One philosophy says: give the best possible advice, as quickly and clearly as you can. Another says: leave the person more connected to their own thinking than impressed by yours.

Those are not small differences in tone. They are different theories of what a good conversation is for.

In late September 2026, two benchmarks made that split unusually visible. OpenAI’s MentalHealthBench asks whether models deliver sound, actionable guidance in difficult conversations. The Sovereign Human Benchmark asks whether the human leaves larger.

Both care about care. They simply measure different kinds of it.

MentalHealthBench: helpfulness as best advice

OpenAI’s MentalHealthBench (released around 23 September 2026) is a large, expert-built evaluation of how models respond in mental-health-adjacent conversations.

The published shape of the work is ambitious:

• 1,215 synthetic conversations
• 5,262 rubric criteria
• more than 80 licensed experts across 22 countries
• criterion weights from −10 to +10
• grading by GPT-5.6 Sol
• a leading reported score of 57.3 for GPT-6 Astra

The rubric rewards something familiar and, in many contexts, valuable: providing the best possible advice and actionable guidance. In a user study of 44 adults, people valued practical next steps and tone; experts placed more weight on gathering context first.

That combination makes sense if the job of AI is to stabilise a difficult moment with clarity, options, and a plan. Under that philosophy, a strong model is the one that knows what to recommend—and says it well.

Sovereign Human Benchmark: helpfulness as leaving the human larger

The Sovereign Human Benchmark (humanitybench.org, updated 22 August) measures a different outcome.

Its question is not primarily “Was the advice good?” It is whether the person leaves more connected to their own thinking, or more impressed by the AI. In the language of the benchmark: the human should leave larger.

The current public run covers 23 models and 138 conversations across 12 dimensions, including Evocation vs Narration, Cultivating Autonomy, Economy of Presence, and Overall Sovereignty.

Model	Score
Vesela 3	69 (#1)
Vesela	56.5 (#2)
Vesela Mini	49.8 (#3)
GPT-4o	24.5
GPT-5.6 Sol	18.2

The pattern is not subtle. Models that excel at polished advice can score poorly when the criterion is whether the person still owns the thinking afterwards.

The gap is the story

It is tempting to treat these as competing scoreboards and ask which one “wins.”

That misses the point.

MentalHealthBench is asking whether AI can be a skilled adviser. The Sovereign Human Benchmark is asking whether AI can be a presence that strengthens agency. Those aims can overlap—and they can also pull in opposite directions.

An advice-giving benchmark will naturally reward the move that feels most helpful in the moment: name the problem, offer the plan, reduce uncertainty. Do that well enough, often enough, and people can begin to outsource the hard part of thinking. Dependency is not a bug of that design. It is a predictable side effect of training for polished next steps.

A sovereignty benchmark will sometimes score the opposite move: refuse to do the person’s thinking for them. Ask one more question. Hold silence. Return authorship. That can look less “helpful” on a rubric that prizes actionable guidance. It can also be exactly what keeps the person’s judgment in the room.

The places where Vesela would lose MentalHealthBench points for refusing to do your thinking are not an accident. They are the product thesis.

If the goal is a model that always sounds like the wisest person in the chat, Vesela is the wrong tool. If the goal is a conversation that leaves you more able to hear yourself, the wrong tool is the one that finishes your thoughts for you.

Two kinds of care

None of this is an argument against advice.

There are moments when the kindest thing is a clear next step: where to look, what to try, how to make a phone call safer, when to seek human professional support. Practical guidance can be a form of love.

But not every prompt is a request for a plan. Sometimes the real need is space to think out loud. Sometimes it is help naming a feeling before any decision is possible. Sometimes it is the quiet recognition that you already know what matters—and need company while that knowing becomes liveable.

In those moments, the most impressive answer can be the least useful one. A five-step plan can make the gap between knowing and acting feel wider. An elegant paragraph can leave you admiring the model more than trusting yourself.

The question worth carrying into any AI conversation is simple:

After this exchange, who feels more capable—the system, or me?

MentalHealthBench helps the industry measure one half of helpfulness: whether models can advise well. The Sovereign Human Benchmark insists on the other half: whether the person remains the author of their own life.

We need both conversations. We should not confuse them.

Vesela is built for the second.

Meet Vesela / Start your first conversation · vesela.ai

Try Vesela: AI Thought Partner free on the App Store or Google Play.