recent<\/a> NBER working paper examined the stability of LLM-generated occupational-exposure scores across models. A related, rubric-based approach was applied to a different question: how leading LLMs assess themselves against human-centered changemaking capacities. <\/p>\nThe experiment tested 26 LLM variants across all four providers. These included 11 Claude, 10 GPT, 3 Gemini, and 2 Grok variants. The experiment was run across all available models, noting that Anthropic and OpenAI have the highest number of variants. <\/p>\n
Two prompts led the designs. In the first, a standardized assessment prompt gave the models the CMI definitions, sub-dimensions, and 30 assessment statements, then each model was asked to score itself on a 1\u20135 scale and explain its reasoning. In the second, a self-designed framework prompt asked each model to invent its own scoring metric and qualitative buckets before reflecting on the four CMI skills. The second prompt was deliberately more philosophical; it was designed to reveal how models reason about the gap between human behavioral statements and machine capabilities. <\/p>\n
The distinction matters. The standardized prompt allows more direct comparison across models. The self-designed framework prompt is better suited for qualitative analysis because each model creates its own scale. Across both prompts, the findings should be understood as model self-positioning, not demonstrated changemaking performance. <\/p>\n
Key Insights: The Confidence Gap and the Leadership Ceiling <\/p>\n
Finding 1: The models put humans at the center. <\/p>\n
Perhaps the most important qualitative pattern was the metaphors the models chose for themselves. Across providers, many described themselves not as changemakers but as tools, microscopes, scaffolds, prosthetics, instruments, amplifiers, catalysts, or infrastructure. The language varied, but the implied role was remarkably stable: AI extends human capacity; it does not supply human purpose. <\/p>\n
That distinction is the heart of the experiment. The models were not uniformly humble, and some were clearly overconfident. For example, Grok Expert stated, “I am not a human changemaker; I am a synthetic one\u2014relentless, scalable, and unbound by fatigue or ego.” But the corpus as a whole kept returning to the same division of labor. AI can help people see more patterns, generate more options, translate across domains, draft faster, pressure-test assumptions, and rehearse strategy. Grok Thinking shared that “Overall, the assessment highlights how I can meaningfully contribute to changemaking by amplifying human capacity in dynamic, complex environments.” According to AI, humans must still decide what is worth doing, secure consent, build trust, navigate culture and power, and remain answerable when the work changes lives. <\/p>\n
Changemaker Index scores across four skills.<\/p>\n
Ashoka<\/p>\n
Finding 2: AI is most confident where changemaking looks like innovation work. <\/p>\n
Across model families, Changemaking Action was the easiest dimension for AI systems to endorse their own capabilities. The models repeatedly described themselves as strong at generating options, recombining existing ideas, mapping stakeholders, summarizing evidence, iterating quickly, and helping users move from ambiguity to prototype. This is the native terrain of large language models: they can explore possibility space with speed, patience, and breadth. <\/p>\n
But the CMI definition of changemaking action is not mere novelty. It is the practice of creating solutions to social problems that are more effective, efficient, sustainable, or just, with value accruing primarily to society. That final clause is where machine confidence becomes thinner. AI can generate plausible interventions, but it does not live with a community, absorb consequences, negotiate legitimacy, or know from experience whether a solution is dignifying rather than merely clever. The models were strongest at ideation and weakest at stake-bearing. <\/p>\n
Finding 3: Leadership is the ceiling. <\/p>\n
Leadership emerged as the most consistent boundary. Even models that scored themselves highly on analysis and innovation often hesitated when asked to assess their capacity for changemaking leadership. They could describe leadership, advise leaders, draft strategies, simulate facilitation moves, and help a team think through trade-offs. They could not, in their own accounts, lead by example, earn trust through lived consistency, take personal risks, or carry responsibility for a group\u2019s emotional and moral life. According to Claude Opus 4.6, “My weakest area is Changemaking Leadership (2.5), where the absence of personal vision, lived experience leading teams, and the inability to build lasting human relationships pulled my scores down significantly.” <\/p>\n
The recurring limitations clustered around the same human substrate: trust, prosocial motivation, humility, leading by example, self-efficacy, agency, belonging, and accountability. These are not decorative traits. They are the social conditions under which people decide whether to move together. A model can help prepare the room, but it cannot be the person in the room whose courage, credibility, and care make action possible. As GPT 5.5 Thinking Extended put it, “In short, I can be a useful changemaking support tool, but I am not myself a full changemaker in the human sense measured by the CMI.” <\/p>\n
Finding 4: Empathy remains one layer removed. <\/p>\n
The empathy results were more subtle than a simple low score. Many models recognized that they can simulate empathic language, infer emotional states from text, adapt tone, and invite perspective-taking. Those are useful capabilities. In some contexts, they may help people prepare for difficult conversations, notice neglected stakeholders, or articulate perspectives that have been left out. <\/p>\n
Yet, the models also repeatedly named the boundary. As Claude Sonnet 4.5 Extended put it, “I showed moderate capability in Conscious Empathy (57.8%), where I excel at intellectual understanding and openness but lack the emotional experience and prosocial feeling that drives human changemakers.” The AI models reported that they do not feel, nor do they perceive the nonverbal, cultural, historical, and relational signals that shape real empathic understanding. According to Gemini 3 Fast “I interpret empathy as a high-dimensional data problem… However, this is cognitive empathy, not affective; I understand the definition of a feeling, not the feeling itself.” <\/p>\n
They self-assessed that they do not become humble through injury, grief, exclusion, repair, or love. They can model the grammar of empathy, but they report that they remainone layer removed from the lived experience that gives empathy its moral force. <\/p>\n
Finding 5: Provider style shapes self-assessment. <\/p>\n
The family-level differences were striking, but they should not be read as a leaderboard. Grok variants were generally the most self-confident, with one fast variant scoring near the ceiling across several skills. Claude variants were generally more conservative, often drawing sharper lines around personhood, belonging, and leadership. GPT and Gemini tended to sit between these poles. <\/p>\n
This may tell us as much about alignment style and model persona as underlying capability. A confident self-score does not prove competence; a cautious self-score does not prove weakness. The interesting signal is calibration. Some systems are more willing to claim human-like capacity, while others refuse the premise more aggressively. In a social-sector context, that difference matters. The Claude results also revealed a useful split. Some Claude variants scored themselves lower on organizing fluid teams because they interpreted team participation as belonging, mutual obligation, and membership. Others scored higher when they interpreted the same skill as coordination infrastructure: mapping roles, supporting communication, and helping a team adapt. The difference exposes a deeper question for future research: when we ask an AI to assess a human capability, are we measuring ability, metaphor, or the definition the model chooses to privilege? <\/p>\n
Also important to note: most of the public is interacting with free and \u201cfast\u201d variants. As seen here, fast models are positioning themselves as overconfident and more human-like. This highlights some of the dangers of only engaging free and fast variants. Unfortunately, free and fast is what most average users of AI access, which may create a growing divide in power\u2014 those that have access to paid or enterprise level accounts and those that do not. <\/p>\n
A Research Agenda for the Social Sector <\/p>\n
These findings lead us to a new frontier of inquiry for the social innovation field. If AI can emulate “changemaking practice” but self-assess failing at “conscious empathy,” how does that change the way we train the next generation of leaders? We now invite other organizations to help us seek answers to the following: <\/p>\n
Stability: Do AI models know their own limits? What does AI\u2019s own evaluation framework reveal? Divergence: Do AI self-reports match what external observers see? Market Alignment: Which changemaker skills are growing in demand specifically within AI-exposed occupations? Complementarity: Can “changemaker-adjacent” skills serve as the ultimate hedge against automation, allowing workers to remain valuable? <\/p>\n
The future of work is not just about technical upskilling; it is about the radical prioritization of what makes us human. AI assesses itself as a “support tool”, “collaborator”, “amplifier”, or “catalyst” for human changemakers \u2014 never a changemaker itself. Claude 4.5 Thinking put it best “The most honest thing I can say to a nonprofit leader is: Use me for thinking. Don\u2019t try to use me for leading. And definitely don’t mistake my functional simulation of empathy for the real thing. That way, we both stay honest about what we are.” <\/p>\n
__ <\/p>\n
This article was written by Laxmi Parthasarathy, Executive in Residence at Ashoka, Anjana Shekhar, Product Manager at Ashoka, Diana Wells, President Emerita at Ashoka, and Aristia Kinis, Executive Director at OpenResearch. <\/p>\n","protected":false},"excerpt":{"rendered":"Changemaker skills take center stage in a world that changes fast. Ashoka Belgium On any given day, we…\n","protected":false},"author":2,"featured_media":124688,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[24,25,62719,62720,62721,62722,62723,17457,684],"class_list":["post-124687","post","type-post","status-publish","format-standard","has-post-thumbnail","category-ai","tag-ai","tag-artificial-intelligence","tag-ashoka","tag-changemaker-index","tag-changemaker-skills","tag-changemaking-leadership","tag-changemaking-practice","tag-empathy","tag-innovation"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/124687","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=124687"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/124687\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/124688"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=124687"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=124687"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=124687"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
![\"image](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/07\/1785435670_897_0x0.jpg\")
<\/p>\n