{
  "$type": "site.standard.document",
  "content": "---\ntitle: \"Seeing AI tasks through a TAM lens\"\ndescription: \"AI adoption research keeps asking 'do you use ChatGPT?' when it should ask\n  'for which tasks?' A task-level framework for thinking about when LLMs\n  actually save you time.\"\ntags:\n  - ai\n---\n\nWhen it comes to AI adoption research, we keep asking \"do you use ChatGPT?\" when\nwe should be adding \"for which specific tasks?\"\n\nThe\n[Technology Acceptance Model](https://en.wikipedia.org/wiki/Technology_acceptance_model)\n(TAM) has been successfully applied to understand LLM\nadoption---[surveys of hundreds of users](https://www.tandfonline.com/doi/full/10.1080/10447318.2024.2314358)\nconfirm that perceived usefulness and ease of use predict whether people adopt\nthese tools. But these studies treat adoption as a technology-level decision\nrather than examining the specific _tasks_ where these things are useful.\n\nA researcher might find an LLM really useful for literature search but useless\nfor theoretical analysis, while a journalist might rely on it for drafting\nroutine updates but avoid it for investigative work. A disinfo actor will\ngladly use it to\n[\"flood the zone\"](https://www.mediamatters.org/steve-bannon/misinformer-year-steve-bannons-flood-zone-shit-approach-destroying-american-democracy),\nbut not for writing to their loved ones. The same person using the same model\nwould give very different answers to \"are LLMs useful\" depending on the task\nat hand.\n\nSo I propose a different framework[^branding]: treat each (user, model, task)\ntriple as its own thing, mapped onto a two-dimensional plane. One axis is\n_familiarity_---how much you've actually used this specific model for this\nspecific task. It ranges from \"never tried\" to \"extensive experience\", and it\nonly moves in one direction: you can't become less familiar with something\nyou've tried.\n\nThe other axis is _usefulness_: whether using the model actually helps you\nget the task done faster. It ranges from actively slowing you down through\nmaking no real difference to genuinely speeding you up. This corresponds to\nTAM's \"perceived usefulness\" but evaluated task-specifically and focused on\ntime/effort rather than moral dimensions[^ethics]. Crucially, this assessment\ncan change as familiarity increases---a tool that initially seems fast might\nprove time-consuming once you factor in error correction, or initial setup\noverhead might give way to genuine productivity gains. And there's always\ngoing to be noise (stochastic parrots and all that) so it's really just an\n\"in general\" judgement.\n\nThe same user with the same AI model might simultaneously occupy multiple zones\ndepending on task, with a substantial neutral zone in the middle: tasks where\nusing the tool takes about as much time and effort as doing it yourself.\nReading the diagram below zone-by-zone fills it out.\n\n<svg class=\"tam-diagram\" width=\"100%\" viewBox=\"0 0 600 500\" xmlns=\"http://www.w3.org/2000/svg\">\n  <!-- Background zones -->\n  <rect x=\"50\" y=\"50\" width=\"250\" height=\"133\" class=\"zone-green-light\"/>\n  <text x=\"175\" y=\"110\" text-anchor=\"middle\" font-size=\"14\" class=\"text-green\" font-weight=\"600\">early positive</text>\n  <text x=\"175\" y=\"128\" text-anchor=\"middle\" font-size=\"12\" class=\"text-green\">impression</text>\n  <rect x=\"300\" y=\"50\" width=\"250\" height=\"133\" class=\"zone-green-dark\"/>\n  <text x=\"425\" y=\"110\" text-anchor=\"middle\" font-size=\"14\" class=\"text-green-dark\" font-weight=\"600\">proven</text>\n  <text x=\"425\" y=\"128\" text-anchor=\"middle\" font-size=\"12\" class=\"text-green-dark\">integration</text>\n  <rect x=\"50\" y=\"183\" width=\"250\" height=\"134\" class=\"zone-orange-light\"/>\n  <text x=\"175\" y=\"243\" text-anchor=\"middle\" font-size=\"14\" class=\"text-orange\" font-weight=\"600\">neutral zone</text>\n  <text x=\"175\" y=\"261\" text-anchor=\"middle\" font-size=\"12\" class=\"text-orange\">(premature/speculative)</text>\n  <rect x=\"300\" y=\"183\" width=\"250\" height=\"134\" class=\"zone-orange-dark\"/>\n  <text x=\"425\" y=\"243\" text-anchor=\"middle\" font-size=\"14\" class=\"text-orange\" font-weight=\"600\">neutral zone</text>\n  <text x=\"425\" y=\"261\" text-anchor=\"middle\" font-size=\"12\" class=\"text-orange\">(informed)</text>\n  <rect x=\"50\" y=\"317\" width=\"250\" height=\"133\" class=\"zone-red-light\"/>\n  <text x=\"175\" y=\"377\" text-anchor=\"middle\" font-size=\"14\" class=\"text-red\" font-weight=\"600\">premature</text>\n  <text x=\"175\" y=\"395\" text-anchor=\"middle\" font-size=\"12\" class=\"text-red\">rejection</text>\n  <rect x=\"300\" y=\"317\" width=\"250\" height=\"133\" class=\"zone-red-dark\"/>\n  <text x=\"425\" y=\"377\" text-anchor=\"middle\" font-size=\"14\" class=\"text-red-dark\" font-weight=\"600\">informed</text>\n  <text x=\"425\" y=\"395\" text-anchor=\"middle\" font-size=\"12\" class=\"text-red-dark\">rejection</text>\n  <!-- Axes -->\n  <line x1=\"50\" y1=\"450\" x2=\"550\" y2=\"450\" class=\"axis\" stroke-width=\"2\" marker-end=\"url(#arrowhead)\"/>\n  <line x1=\"50\" y1=\"450\" x2=\"50\" y2=\"50\" class=\"axis\" stroke-width=\"2\" marker-end=\"url(#arrowhead)\"/>\n  <!-- Arrow markers -->\n  <defs>\n    <marker id=\"arrowhead\" markerWidth=\"10\" markerHeight=\"10\" refX=\"9\" refY=\"3\" orient=\"auto\">\n      <polygon points=\"0 0, 10 3, 0 6\" class=\"arrowhead\"/>\n    </marker>\n  </defs>\n  <!-- Axis labels -->\n  <text x=\"300\" y=\"485\" text-anchor=\"middle\" font-size=\"16\" class=\"axis-label\" font-weight=\"600\">Familiarity →</text>\n  <text x=\"25\" y=\"250\" text-anchor=\"middle\" font-size=\"16\" class=\"axis-label\" font-weight=\"600\" transform=\"rotate(-90, 25, 250)\">Usefulness →</text>\n  <!-- Grid lines -->\n  <line x1=\"300\" y1=\"50\" x2=\"300\" y2=\"450\" class=\"grid-line\" stroke-width=\"1\" stroke-dasharray=\"5,5\"/>\n  <line x1=\"50\" y1=\"183\" x2=\"550\" y2=\"183\" class=\"grid-line\" stroke-width=\"1\" stroke-dasharray=\"5,5\"/>\n  <line x1=\"50\" y1=\"317\" x2=\"550\" y2=\"317\" class=\"grid-line\" stroke-width=\"1\" stroke-dasharray=\"5,5\"/>\n</svg>\n\nThe upper-right quadrant---high familiarity and genuinely speeds you up---is\nthe good place. You've got extensive experience, you've developed effective\nprompting strategies, you understand the limitations, and you've integrated it\ninto your workflow in ways that genuinely save time. This might be generating first drafts\nof routine correspondence that need minimal editing. For me (an academic AI\nresearcher and software developer) things like writing scripts to automate\nrepetitive tasks and generating boilerplate code are in this bucket.\n\nThe middle-and-lower right (high familiarity, but no difference or actively\nslowing you down) is informed rejection. You haven't simply failed to try;\nyou've engaged extensively and concluded the tool doesn't actually save\ntime, or actively wastes it. You've hit the model's limitations repeatedly\nenough to form a stable assessment.\n\nThe neutral zone is interesting: tasks where using the tool takes about as much\ntime and effort as doing it yourself, so you might use it or might not depending\non mood (or whether you're paying for it or you're being\n[subsidised by Silicon Valley VCs](https://www.wheresyoured.at/why-everybody-is-losing-money-on-ai/)).\nA researcher might edit AI-generated literature summaries that need as much\nwork to fix as they would to write from scratch, or a designer might tweak\nAI-generated layouts that never quite capture the intended aesthetic.\n\nThe genuinely slowing-you-down zone includes complex reasoning tasks,\nfact-checking mission-critical information, or creative work requiring genuine\noriginality where the tool actively wastes your time with output that needs\nextensive correction or produces misleading results you have to untangle. In my\nuse of Claude Code (and friends) some coding tasks fall into this for sure,\nadding technical debt that takes longer to fix than writing correctly from\nscratch.\n\nThe upper-left---low familiarity but seems to speed you up---is the early\npositive impression. You've tried it once or twice and it seemed to save time,\nbut you haven't encountered edge cases or the full overhead of error\ncorrection. This zone is unstable: continued use drives rightward along the\nfamiliarity axis, but perceived usefulness might shift up or down as\nexperience accumulates and you discover hidden time costs.\n\nThe middle-and-lower left, finally, is premature rejection or speculative\navoidance. Either you tried it once, found it took just as long or longer, and\ngave up, or you suspect the model can't handle it efficiently based on general\nreputation. Unlike informed rejection, these assessments are speculative\nrather than experiential.\n\nThis framing matters because it cuts against how we usually talk about AI\nadoption. Consider an academic researcher (hey---I write what I know). I might map my\nexperience like this:\n\n- _vibecoding all the things_: upper right---see elsewhere on my blog, but I've\n  overall seen an increase in my productivity, genuinely getting more done\n  faster\n- _literature search and summarisation_: upper right---proven to save time\n  finding relevant papers quickly\n- _drafting standard email responses_: upper right---saves time, works\n  consistently with minimal editing needed\n- _generating first drafts of \"methods sections\"_: middle right---tried\n  extensively, but I end up rewriting so much it takes about the same time as\n  writing from scratch\n- _writing out core/key arguments_: lower right for sure---consistently produces\n  superficial output that takes longer to fix than to write properly in the\n  first place\n\nTechnology-level TAM would aggregate these into a single \"perceived usefulness\"\nscore for \"LLMs in academic research\". But that obscures the actual pattern:\nwhich is much lumpier and task-dependent, with some tasks genuinely saving time\nand others actively wasting it.\n\nThe framework also explains why people's assessments differ so dramatically.\nWe're often arguing about different tasks while thinking we're arguing about the\nsame technology. When someone says \"LLMs are transformative for writing\" and\nsomeone else says \"they're useless\", they might both be right, just talking\nabout different writing tasks with genuinely different time costs.\n\nThis matches what\n[recent meta-analyses](https://www.nature.com/articles/s41562-024-02024-1) have\nfound: human-AI combinations perform significantly differently depending on task\ntype, with performance losses in decision-making tasks but gains in content\ncreation tasks. The tool's utility is entirely task-dependent. Which really\nshouldn't surprise anyone, but here we are.\n\nYou've probably seen people claim AI will \"transform\" work in their domain.\nMost \"transformation\" claims turn out to be one of three things: automation of\ntasks already in the upper-right quadrant (where the tool genuinely speeds you\nup on familiar tasks, and/or can be used to write a script to fully automate a\ntask); magnification of existing patterns, making fast workers slightly faster\nand inefficient workers slightly more inefficient; or wishful thinking about\ntasks currently in the lower quadrants, particularly by folks selling AI\ntools.\n\nReal transformation would mean moving tasks between zones in systematic ways.\nThat might happen as models improve and prompting strategies evolve, but\nit's an empirical question, not a foregone conclusion.\n\nA lot of supposed \"AI-assisted\" work actually lives in the neutral zone:\nusing the tool because it's there and everyone else is, even when it's not\nactually saving time.\n\nSo what do you do with all this? For users: instead of asking \"should I use\nChatGPT?\", ask \"for which specific tasks does it actually save me time after\nsufficient practice?\" This encourages experimentation while validating\ninformed rejection.\n\nDifferent zones need different strategies:\n\n- upper-right tasks deserve workflow integration and continued refinement to\n  maximise time savings\n- middle-right neutral tasks might warrant periodic re-evaluation as models\n  improve, or just accepting they're optional time-wise\n- lower-right time-wasting tasks probably aren't worth more effort unless models\n  substantially improve their speed/accuracy trade-offs\n- upper-left tasks need systematic testing to see if initial time savings hold\n  up under regular use\n- lower-left tasks might benefit from one serious attempt with better prompting\n  before abandoning them\n\nFor researchers: adopt task-level analysis. Ask participants to identify\nspecific tasks, plot them in this space, track how positions change over time\n(evaluate!). This would reveal patterns currently hidden by aggregation. Given\nthat\n[individual performance gains from AI systems depend on task-technology fit](https://aisel.aisnet.org/ecis2020_rp/200/),\nwe need frameworks that capture this task-level variation.\n\nCaveats, of course. All models are wrong, some are useful. Is this one useful? Maybe... I'm still\nthinking it through.\n\nThe two dimensions do miss important factors---confidence, cost, and especially\nethics/responsible use questions[^ethics]. The discrete task framing might\nobscure the fluid, exploratory way people actually interact with LLMs. The\nfamiliarity axis collapses several concepts: frequency of use, diversity of use\ncases, quality of prompting strategies.\n\nBut even a simplified framework beats treating adoption as a binary\ntechnology-level decision. At minimum, it captures the obvious truth that your\nrelationship with these tools is task-specific, and aggregating across tasks\nobscures more than it reveals.\n\nThe real test will be whether thinking in these terms helps you make better\ndecisions about where to invest effort learning these tools. For me, it's been\nclarifying---it legitimises informed rejection (neo-luddism) while encouraging\nstrategic experimentation, and it helps me recognise when I'm in that neutral\nzone where using the tool is more about performance than productivity.\n\nYour mileage may vary.\n\n## Update (March 2026)\n\nAfter running some executive AI training recently we used this as a \"stick\npost-its on the grid\" activity. It was quite illuminating, but my colleague the\nawesome [Lorenn Ruster](https://lorenn.medium.com/) suggested (esp. when running\nit as an activity) that it's a bit easier for folks to get their heads around a\n2x2 grid rather than the 3x2 one above. So, here's the updated version (with\nwhimsical quadrant names):\n\n<svg class=\"tam-diagram\" width=\"100%\" viewBox=\"0 0 600 500\" xmlns=\"http://www.w3.org/2000/svg\">\n  <!-- Background zones -->\n  <rect x=\"50\" y=\"50\" width=\"250\" height=\"200\" class=\"zone-green-light\"/>\n  <text x=\"175\" y=\"145\" text-anchor=\"middle\" font-size=\"14\" class=\"text-green\" font-weight=\"600\">the land of hopes</text>\n  <text x=\"175\" y=\"163\" text-anchor=\"middle\" font-size=\"14\" class=\"text-green\" font-weight=\"600\">& dreams</text>\n  <rect x=\"300\" y=\"50\" width=\"250\" height=\"200\" class=\"zone-green-dark\"/>\n  <text x=\"425\" y=\"145\" text-anchor=\"middle\" font-size=\"14\" class=\"text-green-dark\" font-weight=\"600\">becoming a part</text>\n  <text x=\"425\" y=\"163\" text-anchor=\"middle\" font-size=\"14\" class=\"text-green-dark\" font-weight=\"600\">of BAU</text>\n  <rect x=\"50\" y=\"250\" width=\"250\" height=\"200\" class=\"zone-red-light\"/>\n  <text x=\"175\" y=\"350\" text-anchor=\"middle\" font-size=\"14\" class=\"text-red\" font-weight=\"600\">to investigate</text>\n  <rect x=\"300\" y=\"250\" width=\"250\" height=\"200\" class=\"zone-red-dark\"/>\n  <text x=\"425\" y=\"345\" text-anchor=\"middle\" font-size=\"14\" class=\"text-red-dark\" font-weight=\"600\">a dud</text>\n  <text x=\"425\" y=\"363\" text-anchor=\"middle\" font-size=\"14\" class=\"text-red-dark\" font-weight=\"600\">(for now)</text>\n  <!-- Axes (centered) -->\n  <defs>\n    <marker id=\"arrowhead2\" markerWidth=\"10\" markerHeight=\"10\" refX=\"9\" refY=\"3\" orient=\"auto\">\n      <polygon points=\"0 0, 10 3, 0 6\" class=\"arrowhead\"/>\n    </marker>\n    <marker id=\"arrowhead2-rev\" markerWidth=\"10\" markerHeight=\"10\" refX=\"1\" refY=\"3\" orient=\"auto\">\n      <polygon points=\"10 0, 0 3, 10 6\" class=\"arrowhead\"/>\n    </marker>\n  </defs>\n  <line x1=\"50\" y1=\"250\" x2=\"550\" y2=\"250\" class=\"axis\" stroke-width=\"2\" marker-start=\"url(#arrowhead2-rev)\" marker-end=\"url(#arrowhead2)\"/>\n  <line x1=\"300\" y1=\"450\" x2=\"300\" y2=\"50\" class=\"axis\" stroke-width=\"2\" marker-start=\"url(#arrowhead2-rev)\" marker-end=\"url(#arrowhead2)\"/>\n  <!-- Axis labels -->\n  <text x=\"80\" y=\"270\" font-size=\"13\" class=\"axis-label\">haven't tried</text>\n  <text x=\"520\" y=\"270\" text-anchor=\"end\" font-size=\"13\" class=\"axis-label\">using it all the time</text>\n  <text x=\"320\" y=\"440\" font-size=\"13\" class=\"axis-label\">unhelpful</text>\n  <text x=\"320\" y=\"70\" font-size=\"13\" class=\"axis-label\">helpful</text>\n</svg>\n\n[^branding]: The framework doesn't have a name, because I'm bad at branding... hmm.\n\n[^ethics]:\n    The moral/ethical dimension matters enormously---things like\n    [digital phrenology](https://gizmodo.com/were-doing-ai-phrenology-again-2000553600)/[racism](https://hai.stanford.edu/news/covert-racism-ai-how-language-models-are-reinforcing-outdated-stereotypes)\n    and racism aren't just \"not useful\", they're harmful. A tool can be both\n    fast and unethical, or slow and ethical. The framework deliberately focuses\n    more on the pragmatic \"does this save me time?\" dimension while\n    acknowledging that ethics is a separate consideration. If you like, add a\n    third \"is it good for human flourishing\" dimension.\n\n<!-- styles for the TAM diagram SVG (light/dark mode) -->\n<style scoped>\n.tam-diagram .zone-green-light { fill: #e8f5e9; opacity: 0.5; }\n.tam-diagram .zone-green-dark { fill: #a5d6a7; opacity: 0.7; }\n.tam-diagram .text-green { fill: #2e7d32; }\n.tam-diagram .text-green-dark { fill: #1b5e20; }\n.tam-diagram .zone-orange-light { fill: #fff3e0; opacity: 0.5; }\n.tam-diagram .zone-orange-dark { fill: #ffe0b2; opacity: 0.7; }\n.tam-diagram .text-orange { fill: #e65100; }\n.tam-diagram .zone-red-light { fill: #ffebee; opacity: 0.5; }\n.tam-diagram .zone-red-dark { fill: #ef9a9a; opacity: 0.7; }\n.tam-diagram .text-red { fill: #c62828; }\n.tam-diagram .text-red-dark { fill: #b71c1c; }\n.tam-diagram .axis { stroke: #333; }\n.tam-diagram .arrowhead { fill: #333; }\n.tam-diagram .axis-label { fill: #333; }\n.tam-diagram .grid-line { stroke: #999; }\n\n.dark .tam-diagram .zone-green-light { fill: #1b5e20; opacity: 0.4; }\n.dark .tam-diagram .zone-green-dark { fill: #2e7d32; opacity: 0.5; }\n.dark .tam-diagram .text-green { fill: #a5d6a7; }\n.dark .tam-diagram .text-green-dark { fill: #c8e6c9; }\n.dark .tam-diagram .zone-orange-light { fill: #e65100; opacity: 0.3; }\n.dark .tam-diagram .zone-orange-dark { fill: #e65100; opacity: 0.45; }\n.dark .tam-diagram .text-orange { fill: #ffcc80; }\n.dark .tam-diagram .zone-red-light { fill: #b71c1c; opacity: 0.35; }\n.dark .tam-diagram .zone-red-dark { fill: #c62828; opacity: 0.5; }\n.dark .tam-diagram .text-red { fill: #ef9a9a; }\n.dark .tam-diagram .text-red-dark { fill: #ffcdd2; }\n.dark .tam-diagram .axis { stroke: #ccc; }\n.dark .tam-diagram .arrowhead { fill: #ccc; }\n.dark .tam-diagram .axis-label { fill: #ccc; }\n.dark .tam-diagram .grid-line { stroke: #666; }\n</style>\n",
  "createdAt": "2026-05-13T23:14:43.134Z",
  "description": "AI adoption research keeps asking 'do you use ChatGPT?' when it should ask 'for which tasks?' A task-level framework for thinking about when LLMs actually save you time.",
  "path": "/blog/2025/11/12/seeing-ai-tasks-through-a-tam-lens",
  "publishedAt": "2025-11-12T00:00:00.000Z",
  "site": "at://did:plc:tevykrhi4kibtsipzci76d76/site.standard.publication/self",
  "tags": [
    "ai"
  ],
  "textContent": "AI adoption research keeps asking 'do you use ChatGPT?' when it should ask 'for which tasks?' A task-level framework for thinking about when LLMs actually save you time.",
  "title": "Seeing AI tasks through a TAM lens"
}