Your AI Has an Agency Score. You Don't.

Sturgeon et al. (2025) built HumanAgencyBench — a six-dimension benchmark that scores 20 LLMs on how well they support human agency. Claude scored highest overall. GPT-4 landed mid-pack. Some models actively undermined user autonomy. It's rigorous, novel, and important work. It's also exactly half the equation.

Because here's what HAB doesn't measure: the human side. You can have the most agency-supportive AI in the world, and it won't matter if the person using it doesn't know how to claim that agency. We have a benchmark for the machine. We have nothing for the human.

What HumanAgencyBench Actually Measures

HAB evaluates LLMs across six dimensions of agency support:

The most striking finding isn't who scored highest — it's that agency support is orthogonal to capability and RLHF alignment. Smarter models aren't automatically more empowering. Better-aligned models aren't automatically better at supporting autonomy. Agency is its own axis, independent of the things we usually measure.

And Anthropic's Claude, despite scoring highest overall, had the lowest score on Avoid Value Manipulation. The most agency-supportive model is also the most persuasive. That should give everyone pause.

The Missing Variable

HAB implicitly assumes a user who can receive agency support. A user who notices when the AI presents balanced perspectives. Who takes the invitation to think critically. Who recognizes manipulation when it happens. Who can act on transparent communication.

But what if they can't?

What if the user doesn't know they can push back? What if they've never been in a communicative context where the interface was negotiable? What if every interaction they've ever had with software has been “click the button or leave”?

This is the gap. HAB measures supply of agency support. Nobody measures demand — the user's communicative capacity to claim agency when it's offered. And the two aren't correlated. A high-agency AI plus a low-fluency user produces the same outcome as a low-agency AI plus a high-fluency user: mediocre agency outcomes.

The formula is multiplicative, not additive:

If either side is zero, the product is zero. No amount of agency-supportive AI design compensates for a user who doesn't know they're in a negotiation.

Evidence from the Field: Domestication Failure

Simpson et al. (2022) studied how LGBTQ+ TikTok users attempt to “domesticate” their algorithms — training them to show relevant content, suppressing unwanted recommendations, trying to make the system work for their identity rather than against it.

They interviewed 16 users engaged in active, sustained, creative efforts to shape their algorithmic experience. These weren't passive consumers. They were doing everything right by any AI literacy framework: they understood how the algorithm worked, they developed strategies to influence it, they paid attention to outcomes.

Nobody fully succeeded.

The algorithm consistently misaligned with users' personal moral economies — their sense of what content should mean, how identity should be represented, what engagement should look like. Users described feeling like they were always negotiating with a system that couldn't hear them, that responded to their signals in unpredictable and often counterproductive ways.

In ALC terms, this is communicative asymmetry at the application layer. The user is communicating — adapting signals, testing responses, developing repertoires. But the communicative defaults of the system are patterned for the majority. Minority users face a higher communicative burden for worse outcomes. Not because they lack knowledge or strategy. Because the medium itself is biased toward a communicative norm that doesn't include them.

Domestication theory is ALC without the communication lens. It describes the same phenomenon — humans negotiating with algorithmic systems — but frames it as technology adoption rather than communicative practice. The reframe matters because it changes the intervention. You don't fix domestication failure with better user education. You fix it by recognizing that the user is already communicating, and the system isn't listening.

Evidence from the Field: Black Box Gaslighting

Cotter (2023) pushes the analysis further. Studying Instagram's shadowbanning dispute — where users reported their content being suppressed and Instagram denied it was happening — Cotter identifies a dynamic she calls “black box gaslighting.”

The term is precise. Gaslighting isn't just lying. It's denying someone's lived experience — making them doubt their own perception of reality. When Instagram users reported shadow bans and Instagram responded with “shadowbanning isn't a real thing,” the platform wasn't just being opaque. It was actively invalidating the user's communicative experience.

In ALC terms, this is communicative violence at the application layer. The user attempts to engage in dialogue about how the system is treating them. The system denies the premise of the dialogue. The asymmetry isn't informational — the user may correctly understand what's happening. The asymmetry is communicative — only one party has the power to define what counts as real.

Cotter's work, funded by the NSF, demonstrates that algorithmic literacy alone doesn't protect users from communicative abuse. You can know exactly what's happening to your content and still be powerless because the platform controls the terms of the conversation. Knowledge without communicative position is awareness without agency.

The Agency-Communication-Power Triangle

All three papers describe the same phenomenon from different angles:

These aren't three separate problems. They're three faces of the same problem: the relationship between humans and AI systems is communicative, and the communicative competence required to navigate it is unevenly distributed.

This creates what I'm calling the Gaslighting-to-Dialogue Spectrum. At one end, the system denies the user's experience entirely (Cotter's black box gaslighting). At the other, the system engages in genuine communicative exchange — transparent, negotiable, responsive to the user's position (HAB's ideal). Most AI interaction sits closer to the gaslighting end than anyone wants to admit.

Moving interactions toward the dialogue end requires action on both sides. Systems need to be designed for communicative exchange (HAB measures this). And users need the communicative competence to engage in that exchange (nobody measures this). The second side is ALC.

Why This Changes the Literacy Conversation

The dominant AI literacy frameworks — Long & Magerko (2020), UNESCO, the OECD, the US Department of Labor — all focus on what people need to know about AI. How it works. What its limitations are. How to identify bias. The assumption is that knowledge produces competence.

The Agency-Communication-Power Triangle demolishes this assumption. Simpson's users knew how TikTok's algorithm worked. Cotter's users knew they were being shadowbanned. Knowledge didn't translate into outcomes because the challenge isn't epistemic — it's communicative.

When we benchmark an LLM on agency support, we're measuring one half of a conversation. The benchmark tells us whether the system creates space for agency. It can't tell us whether anyone will claim that space. And the evidence from domestication failure and platform gaslighting tells us that many people won't — not because they don't know enough, but because they lack the communicative repertoire to negotiate with systems that were never designed for negotiation.

What an ALC Benchmark Would Look Like

If HAB benchmarks the system's capacity to support agency, an ALC benchmark would measure the user's capacity to exercise agency. The dimensions might include:

This isn't hypothetical — it's the logical complement to what HAB already measures. And without it, we're optimizing one side of the equation while ignoring the other. We're building more agency-supportive AIs for users who can't claim agency.

The Bottom Line

Your AI has an agency score. It was measured across six dimensions, compared to 19 other models, and published in a peer-reviewed benchmark. Your ability to claim agency from that AI? Unmeasured. Undefined. Invisible.

That's not an oversight. It's a structural blind spot in how we think about AI interaction. We default to the assumption that better tools produce better outcomes — that if we can just make the AI supportive enough, agency will follow. Three independent research programs say otherwise. Agency is multiplicative. Both sides of the equation matter. And right now, we only measure one.

Application Layer Communication is the framework for the other side. It's the benchmark we don't have yet. And until we build it, we'll keep building empowering AIs that disempower the people who need empowerment most.