Drawing the line on AI in social work

An invitation for discussion

The critiques of AI are familiar by now. Large language models hallucinate. They flatter users rather than challenge them (sycophancy). They were trained on scraped work without consent. They consume water and electricity at scale. They sit atop exploited annotation labor. They route private data through the servers of massive corporations with little meaningful oversight. They encode and amplify disparities that social work has spent decades trying to name.

Each of these points contains something true. I have written about several of them myself. I am not asking anyone in the field to pretend the harms are not real.

I want to open a different question, and I have not yet seen the field engage it directly.

If these are the standards you are putting forward, are they fixed standards? Is the position a complete moratorium on AI in social work practice, research, and education? Or are there conditions under which use is acceptable, and if so, what are those conditions, and who gets to decide?

I ask because the current discourse collapses two very different things. One is an ideological stance against a category of technology. The other is a strategic framework for deciding when, how, and for whom a tool is appropriate. These are not the same, and the field needs to be honest about which one is actually on the table.

Consider four cases. I want to be plain about something before I lay them out: these are not hypotheticals constructed to win an argument. All four are drawn from work I am actively doing or have done, with real collaborators and real populations. I raise them because I have not found a principled way to resolve them inside the concerns colleagues are currently raising, and I would rather surface the tension than pretend it does not exist.

Case one: the asylum file that no one has time to write

Asylum seekers in the United States are among the most vulnerable people any social worker will encounter. The paperwork is dense, the evidentiary standards are exacting, and the volume of need dwarfs the supply of attorneys, accredited representatives, and social workers available to help. People with legitimate claims are denied because no one had the capacity to thoroughly document country conditions, organize the narrative, or assemble the exhibits before the filing deadline.

Suppose a qualified representative uses a large language model as a drafting aid. The representative verifies every factual claim and every citation against primary sources, reviews the country conditions evidence for accuracy, and takes full professional responsibility for the final filing. The tool accelerates the work; it does not replace judgment. The result is a stronger, more complete draft in a fraction of the time, produced under the same standard of care the representative would apply to anything leaving the office.

Is using it a violation of the standards you have set out? If the answer is yes, what is your response to the person whose claim goes unfiled because the alternative was nothing? If the answer is no, then the moratorium is not actually a moratorium, and we should say so plainly.

Case two: research that would otherwise not happen

There are research questions in social work that require processing tens of thousands of administrative records, case narratives, or published abstracts. Done manually, the work takes months or years. Much of it never gets done at all, which means the knowledge that would protect children, inform substance use treatment, or evaluate policy effects simply does not exist.

A well-designed NLP pipeline can complete the same work in days, under conditions that make the results defensible: validation against hand-coded gold standard samples, reported error rates, sensitivity analyses for construct drift, and domain-expert review of outputs before anything enters the analytic record. The energy cost is real. It is also a small fraction of the cost of a year of human labor, travel, server time, and institutional overhead to do the work the slow way, assuming the slow way is even possible. More often than not, the realistic alternative is not slower research. It is no research.

Under your standards, is this acceptable? If not, what is the plan for the questions that will go unanswered? If yes, what are the conditions: open-weight models only, local deployment only, documented energy accounting, public release of the pipeline, something else?

Case three: AI that predates generative AI

Many of the tools social work researchers and practitioners have used for years are AI in the technical sense: supervised classifiers, topic models, word embeddings, clustering algorithms, and rule-based NLP. They run locally on a laptop. They do not flatter the user. They do not scrape the internet. They do not consume meaningful water. They were built by researchers using published methods.

Are these also prohibited? If the concern is generative models specifically, the critique needs to say so. If the concern is the word “AI,” then the critique is about a label rather than a practice, and much of the field has been inside the boundary for a long time without anyone noticing.

Case four: a small model running on your laptop

Suppose the model is small enough to run entirely on your own computer, with the network cable unplugged and the wifi turned off. Nothing touches a cloud server. No query is logged by a corporation. No data is ever exposed to a third party. Suppose it was distilled from a larger model, which means most of the training cost has already been paid, and the efficient version inherits that capability with a fraction of the energy footprint during inference. Suppose it handles a specific, bounded task: summarizing a case narrative, drafting a first pass at an outreach letter, classifying records against a codebook.

No data leaves the laptop. No API calls, not even occasional ones. No water consumed per query. The user controls the weights and can audit the behavior. The marginal environmental cost of running it is closer to that of running a spreadsheet than to that of hitting a cloud API. A client’s information stays within the room where it was collected.

Is this acceptable? If the objection is that the small model inherits from a larger one with scraped training data, that history is real, but the same logic applies to every search engine, every compiler, and every piece of software built on public code. If the objection is environmental, the local offline model is the most defensible version of the technology that currently exists. If the objection is data privacy, offline operation directly addresses it. If the answer is still no, then the category being rejected is broader than generative AI, and the field should say so plainly.

The AI is already in the room

The field already operates inside AI systems. GPS routing, spam filters, autocomplete, photo search, fraud detection, voice transcription, medical imaging triage; the infrastructure of ordinary professional life has absorbed machine learning without a moratorium and largely without debate. I note this not to accuse anyone of inconsistency, but to establish the baseline. The question is not whether to participate in AI. That boundary was crossed years ago. The question is which deliberate additions the field will engage with, and on what grounds.

The deskilling concern

A related concern deserves direct treatment: that routine use of these tools will erode the capacities of the practitioners who rely on them. I take the concern seriously, and I think it is misdirected. Deskilling happens when we stop teaching skills. A student who never learns to read a country conditions report, trace a citation, or evaluate a source will be deskilled regardless of what tools are in the room. A student who learns those skills first, and then learns to use tools under supervision, is not. The question is not whether AI is present in the workflow. It is whether the curriculum continues to develop the underlying capacities that let a practitioner supervise the tool rather than defer to it. That is a pedagogical responsibility, and it is ours. Framing deskilling as a property of the technology lets the curriculum off the hook for a problem that belongs to the curriculum.

The conference we flew to

The field has already worked out, for air travel, exactly the kind of framework I am asking for here. Every year, social work researchers fly on fossil fuels to conferences, including conferences where emerging technology ethics are on the program. We do not treat the practice as an ethical crisis, nor as ethically neutral. We weigh purpose against cost, reduce harm at the margins through shared travel, virtual options, fewer trips per year, and carbon accounting in program notes, and we accept that some trips are worth taking and others are not.

That is the reasoning structure I am asking the field to apply to compute. Not blanket acceptance, and not blanket refusal. A clear accounting of which uses are worth which costs, under what conditions, with what mitigations. The framework is mature in one domain, and it translates directly to the other.

The positionality question

There is a version of this critique that functions less as a standard and more as a personal exemption. An established researcher with tenure, a mature research program, and a professional network built over decades can plausibly continue the rest of their career without ever opening a language model. The workflow is intact. The citations still flow. The pipeline was built in a different era and still runs. A career can be built and maintained on those terms, and I know several people doing exactly that.

That is a defensible personal choice. It is not a field-wide standard, and it should not be presented as one.

The students I teach do not have that option. The practitioners inside organizations that are already integrating these tools, with or without their input, do not have that option. The self-represented litigant who cannot afford an attorney does not have that option. The asylum seeker whose case is sitting in a pile does not have that option.

If your ethical framework works for you because your career was already built, the field still needs to reckon with the fact that it does not work for the people whose careers are being built right now, inside a profession that is being reshaped, whether social work engages thoughtfully or not. Knowledge work is moving into an AI-infused environment. That is not a prediction. It is the current operating reality across every adjacent field, and social work will be no exception.

Opting out personally is a choice available to some. It is not an answer to the question the field is facing. It is a way to avoid answering it.

What I am actually asking

I am not asking anyone to drop their concerns. I am asking for specificity.

If the answer is that no use is acceptable, I want to understand how that position holds up against the asylum, research, classifier, and local model cases. What happens to the people and the questions on the other side of that line?

If the answer is that some uses are acceptable, then the work begins. What are the conditions? Open weights? Local deployment? Energy budgets? Human review requirements? Sycophancy mitigation in supervisory and ethics-consultation contexts through prompt design, model selection, and workflow constraints that preserve the challenge function? Population-specific carve-outs? Disclosure standards that go beyond the current boilerplate? A tiered framework that distinguishes research from practice from administration? Who decides, and how do disagreements get adjudicated?

A third option exists in principle: case-by-case review under institutional ethics processes, with the burden on the proposer and no pre-specified decision rule. I would welcome that framework. It is not the framework currently being advanced, and until it is, the conversation tends to settle into either blanket objection or unspecified conditions. Either of the first two can be argued, refined, and held accountable together. A blanket position is harder to work with, because the moment it encounters a hard case, it either bends or it leaves the people the case was about without a plan.

Every critic of generative AI who publishes in a commercial journal is already using a tool built inside an exploitative system. The field’s response to academic publishing has not been refusal; it has been naming the harm, pushing for open access (while recognizing the unanticipated problems that reform has introduced, from predatory journals to article processing charges that disadvantage unfunded scholars), and continuing to use the infrastructure while working to change it. That is the posture I am asking us to bring to AI.

I am open to being convinced that the line should sit in a particular place, and I am open to revising my own current uses in light of a better argument. What I am not able to work with is a position that refuses to draw a line at all, because the people in the cases above cannot wait for one that never gets drawn.