In a hurry? The short version
There is a quiet anxiety running through the therapy world about artificial intelligence. Much of it is justified and concerns consent, data, and the integrity of the therapeutic relationship. But a surprising amount of it has attached itself to a single, vivid worry: the environmental cost. The image of a humming data centre drinking electricity and water every time we ask a model to tidy up a clinical note has become a moral talking point in supervision groups and peer forums.
It is worth replacing the image with numbers. When you do, something useful happens: the environmental question shrinks to roughly its true size, and the harder, more important ethical questions come back into focus.
This piece works through four things, in order. First, what "using AI" actually means in a practice like mine and what each part really costs. Second, the comparison clinicians keep asking about: is a do-it-yourself setup greener than a commercial scribe like Heidi? Third, the question that should have been first all along: what does the client actually get out of it, and does that justify the cost? And finally, the question underneath all of it, the one I find genuinely unsettling: not what this technology costs, but how little, and what that might mean for therapy itself.
The three jobs, and why they're not the same
In a typical week I might ask AI to do three quite different things, and lumping them together is where most of the confusion starts.
Transcription turns recorded or ambient session audio into text. For someone seeing clients five or so hours a day, this is the single largest slice of usage by volume: five hours of speech is a great deal of audio to process.
Summarisation takes those transcripts and produces something useful: a structured clinical note, a letter, a referral, a progress summary for the client.
In-session generative help is different again, and more ethically loaded. This is using a model live in the room: to suggest a cognitive interweave during EMDR processing when a client is looping, or to draft an EMDR narrative or resource-installation story tailored to the work in front of you. It is occasional, it is creative rather than mechanical, and as we'll see, it is energetically trivial but clinically the most sensitive of the three.
The actual energy cost
Start with transcription, because it dominates. A 2024 measurement of OpenAI's Whisper large-v3, the heaviest of the open transcription models, found it consumes roughly 32 watt-hours per hour of audio when run efficiently in batches, generating something like eight grams of CO₂ per hour.1 Five hours of sessions therefore comes to about 160 watt-hours a day, and that figure is on the conservative side: optimised production pipelines (quantised models, batching, the kind of thing a cloud provider runs at scale) routinely beat it by a wide margin.
Summarisation is lighter than people expect. Independent and industry estimates through 2025 have converged on roughly 0.3 watt-hours for a typical text query,2,3 about the energy of running a microwave for a second or two. Session summaries are heavier than a typical query because the transcript going in is long, which pushes each one into the low single-digit watt-hours. Across five or six sessions, that's somewhere around 15 to 30 watt-hours a day.
The in-session generative work (interweaves, narrative stories) is the part clinicians instinctively worry about most and the part that matters least environmentally. An interweave suggestion is a short query; a tailored narrative is a longer one, perhaps one to three watt-hours. Even on a heavy trauma-processing day with a dozen or more such generations, you are adding a few tens of watt-hours at most. It disappears into the rounding.
Put together, a full clinical day lands at roughly 0.2 kilowatt-hours. To make that concrete: it is about the same as simply having your laptop switched on through those five hours of sessions anyway. The AI is, in energy terms, roughly invisible against the device you're already running.
Scaled across a working year of around 230 clinical days, that's about 40 kilowatt-hours, which a domestic fridge-freezer gets through in about six to eight weeks. On the present UK grid (now fairly low-carbon, well under 0.2 kg CO₂ per kWh)4 the annual emissions come to somewhere in the region of six to eight kilograms of CO₂. That is comparable to driving twenty-five to thirty miles, or about half a kilogram of beef. For a sense of proportion against the rest of a life: a single return short-haul flight emits several hundred kilograms, so an entire year of AI-assisted documentation is well under two per cent of one holiday.
Water, the other figure people raise, deserves the same treatment: numbers and comparators rather than imagery. Google's own disclosure puts a median text prompt at about a quarter of a millilitre of water;3 scaled to the usage above, a full clinical day's cooling water lands somewhere under half a litre, and a couple of litres on the most pessimistic accounting that includes the water used to generate the electricity. For scale, that is less than a single toilet flush, and around one per cent of the roughly 137 litres an average person in England and Wales uses directly each day.5 The honest caveat is location: the same litres matter far more in a water-stressed region of Arizona or Spain than in the UK or northern Europe, so if water is your concern, the useful question is where your provider's data centres sit, not whether to abstain. It also follows, absurdly but arithmetically, that a vegetarian therapist who skips one toilet flush a week can run AI all year and still come out greener than an abstaining colleague with a burger habit. Carbon accounting has no respect for moral aesthetics.
None of this counts the one-off cost of training the underlying models. That's a real and large number, but it is shared across many millions of users, so an individual practitioner's marginal slice of it is negligible. The figures above are the cost of use, which is what your decision actually controls. One honest limit on that framing, though: "my marginal slice is negligible" is the same arithmetic that excuses every collective-action problem, and the very cheapness that makes each query trivial is what is driving the worldwide boom in data centres. At the level of one practice, the numbers above hold. At the level of the whole sector and beyond, the aggregate is a legitimate thing to care about; it is just not a reason for one clinician's notes to carry the guilt for it. I come back to the aggregate properly near the end of this piece, because it deserves more than a parenthesis.
DIY versus Heidi: the comparison that misleads
The instinct among privacy-minded, environmentally-minded clinicians is that rolling your own (a self-hosted Whisper instance feeding the Claude API, say) must be the greener and cleaner option than handing sessions to a commercial scribe like Heidi. On the environmental axis, this is usually backwards.
A commercial provider runs transcription on highly utilised hardware in a cooling-optimised data centre, processing thousands of hours back to back. A self-hosted model running on a modest cloud box for one practitioner's five daily hours sits idle most of the time and is cooled less efficiently. Per session, the optimised cloud service is frequently more energy-efficient, not less. "DIY is greener" is, in most realistic setups, a comforting myth.
Which is rather freeing, because it means the genuine DIY-versus-commercial decision was never really about carbon. It is about data governance, and there the trade-offs are real and run the other way:
So choose your tool on the data questions. Treat the energy difference as the tie-breaker it isn't.
The ethics that actually deserve the worry
If the carbon is small, the ethical weight has to go somewhere more honest. Four places, in roughly descending order of importance.
Consent and transparency. Therapy is an unusually intimate setting, and EMDR trauma work especially so. A client has a right to know, in plain language, that a session is being recorded or processed by AI, what happens to that recording afterwards, and that they can decline without it costing them anything in the relationship or the care. Consent here should be specific and revocable, not a clause buried in an intake form. And it is worth being honest that consent in therapy is never quite the free choice a form implies: clients want to please the person they depend on, so a therapist's visible enthusiasm leans on the scales however carefully the question is put. The safest posture is opt-in, asked once, with genuine indifference to the answer. For some trauma clients, the very sense of being recorded can be activating, and that has to be held clinically, not waved through.
Data sensitivity. Session content is among the most sensitive personal data there is. The whole DIY-versus-commercial question above is, properly understood, a data-protection question wearing an environmental costume. Where the audio lives, who can reach it, how long it persists, and whether it trains anything are the things that determine whether this is ethical, and they need answering before a single session is processed. For special-category health data, answering them formally, with a data protection impact assessment rather than a vibes check, is what UK GDPR expects.
Accuracy and the clinical record. Transcription models make mistakes, and not only innocent ones: speech-to-text systems have been documented occasionally fabricating content outright, inserting sentences nobody said, particularly across silences.6 An AI-drafted note that a clinician signs without properly checking becomes a clinical record, with medico-legal weight, containing an error nobody made on purpose. And automation bias is real: the more often the draft is right, the less carefully we read it. Whatever the workflow, the note is yours. Reviewing it against your own memory of the session is not an optional courtesy; it is part of the work, and it is a cost that should be counted against the time the tool saves.
The in-session generative line. Using a model live to suggest an interweave or draft a narrative is categorically different from using one to write up notes afterwards, and it deserves its own caution. Two risks stand out. The first is clinical: the output is a prompt for the therapist's judgement, never a script to read out. The attunement, the timing, the read of this client in this moment remains entirely the clinician's, and there is a genuine de-skilling risk if the tool quietly becomes the source of the clinical move rather than an aid to it. The second is confidential: generating something genuinely tailored often means feeding client material to a model mid-session, which drags the data-governance question right into the live work. Keeping client-identifying detail out of third-party calls, or ensuring the contract explicitly covers it, matters far more here than in batch note-writing.
There's a meta-point worth naming too. Environmental guilt can quietly become a proxy: a respectable-sounding reason to avoid AI when the real hesitation is something else, or, just as unhelpfully, a way to wave away the data ethics by pointing at the reassuringly tiny carbon figure. The two axes are separate. The carbon is small; the consent and data questions are not. Spend the ethical attention where it's actually needed.
What the client gets: the other side of the ledger
Cost is only half a judgement. The right question is cost against benefit, and to whom, and the benefit side is more substantial than the environmental framing tends to allow. But here I should hold myself to the standard I set earlier: the costs above came with measurements, and the benefits below mostly do not. There are, as yet, no good trials showing that AI session summaries improve therapy outcomes. What follows is clinical observation and client feedback, which is evidence of a softer kind, and I would rather say so plainly than smuggle it past you dressed as data.
For the clinician, the gain is real: less time lost to documentation, less of the administrative load that drives burnout, and, not trivially, more presence in the room when you're not half-attending to your own note-taking. A more present therapist is a clinical good, not merely a convenience.
For the client, the benefits are concrete but emphatically not universal:
But for other clients, a written record is the opposite of welcome: a safety concern about who might see it, or a discomfort with anything AI-touched, or a feeling that a summary flattens something that mattered. Benefit has to be offered and individualised, never defaulted on. The summary a client asked for and values is a gift; the same summary produced without their say-so is a breach, and where consent is concerned there are no small ones.
The question underneath the question: what happens when this gets good
A note for any client reading this: the section that follows is me thinking aloud with colleagues about the profession's future. Your therapy, here and now, is with a human, and nothing below changes that.
Everything above treats AI as an administrative assistant. It transcribes, it summarises, it occasionally suggests. The honest next thought is harder, and it is the one I actually lose sleep over. The numbers that make the environmental worry evaporate, fractions of a penny per query, are the same numbers that should make the profession sit up. They describe a technology whose marginal cost is heading towards zero while its capability keeps climbing. The real question was never whether AI costs too much. It is what happens because it costs almost nothing.
I can see a plausible future, and not a distant one, in which a great deal of genuinely high-value therapeutic work is delivered by AI under clinical supervision: structured protocols, psychoeducation, between-session support, and in time substantial parts of the therapeutic conversation itself, with a clinician like me overseeing the work, holding the risk, the formulation, and the moments that should never be left to a machine. In that model the therapist's job changes shape. It becomes less about personally delivering every minute of care and more about supervising care at scale, the way a consultant oversees a service rather than sitting in every appointment.
I should name the tension, because a careful reader will spot it: most of this article reassures you that AI is an administrative aid and the clinician remains the clinician, and this section contemplates the machine moving into the conversation itself. Both can be honest; one describes the present, the other a possible future, and pretending the second cannot happen because the first is comfortable would be the kind of avoidance we charge our clients to notice. Two further cautions belong here. The evidence that AI can deliver therapy, as opposed to psychoeducation and structured exercises, does not currently exist, though "does not exist" needs careful reading: it means untested, not tested and found wanting. A clinical trial takes years to design, run, and publish, and the models improve in months, so any study of chatbot therapy is examining a technology that is already out of date by the time it reaches print. The evidence will permanently lag the capability, and that cuts both ways: we cannot claim AI therapy works, and we also cannot lean on the research silence as if it were reassurance. What the research is silent about, decades of work on human therapy is not: much of therapy's effect lives in the relationship itself, and nobody yet knows whether an alliance with a system for which you do not exist can carry that weight. And the consultant analogy has a known flaw: consultants supervise juniors who become the next consultants. If AI takes the formative work, the profession will have to build its training ladder deliberately, because the market will not preserve it for us.
I will also offer one piece of softer evidence of my own, under the same health warning I gave earlier about anecdote. I recently built a prototype AI interviewer for the assessment side of my own work: the structured, adaptive history-taking that precedes therapy, asking follow-up questions the way a clinician would, with everything it gathered reviewed by me before it counted for anything. Assessment is not therapy; taking a history is not holding a rupture. But what struck me was not how far away the technology felt. It was how close. The distance between "clearly impossible" and "plausibly imminent" closed further in one weekend of building than in years of reading opinion pieces, mine included, and it is why I treat the timeline in this section as a planning question rather than science fiction.
There is a second force pushing the same way, and it comes from outside the consulting room. It helps to be clear about where the UK already stands. Unemployment has climbed to around five per cent, a five-year high, after rising faster over the past year than in any other G7 economy: roughly 300,000 more people out of work in twelve months.7 The OECD expects it to keep climbing through 2026, again the largest rise in the G7,8 and the CBI has warned of unemployment heading towards two million.9 The contrast with our neighbours is striking. While Britain's rate has been climbing, euro-area unemployment has been falling to record lows, 6.1 per cent at the start of 2026, with Germany and the Netherlands down at four.10 The UK's headline rate still sits just below the euro-area average, but the direction of travel is exactly opposite: Europe's labour market has been strengthening while ours weakens, and on the OECD's numbers no comparable economy is deteriorating faster.
Look closer and the pattern has an AI-shaped edge. UK entry-level vacancies have fallen by almost a third since ChatGPT launched,11 and graduate openings are at their lowest in seven years.12 Vacancies in the occupations most exposed to AI have fallen by 37 per cent since late 2022, against 26 per cent elsewhere.13 Youth unemployment is around sixteen per cent, its highest in a decade and above its pandemic peak: roughly one in six young people looking for work.14 None of this proves AI is the sole cause; higher employer National Insurance and minimum-wage costs are doing real work in those numbers too. But the jobs disappearing fastest are precisely the ones AI does most cheaply, and that is the signature you would expect to see first.
Why Britain first, though? If AI were the whole story, Germany would be suffering too. The answer is that two things landed on the same jobs at the same time. In April 2025 the government raised employer National Insurance from 13.8 to 15 per cent and, more significantly, cut the threshold at which it starts from £9,100 to £5,000, making part-time and junior staff sharply more expensive; the minimum wage for 18-to-20-year-olds rose 16.3 per cent the same month, with another 8.5 per cent the following April.15 That raised the price of exactly the labour AI substitutes for, entry-level and routine cognitive work, at exactly the moment the substitute became almost free, and researchers observe firms reaching for AI precisely as their response to rising employment costs.16 Three structural facts then explain why the squeeze shows here before the continent. The UK is overwhelmingly a services economy, concentrated in the white-collar cognitive work that generative AI does best, and in English, the language the models are best at; Germany and Italy carry far larger manufacturing shares that AI cannot yet touch. The UK's famously flexible labour market transmits shocks into unemployment within months, where continental employment protection and short-time-work schemes slow the same adjustment by years. And the euro area's record lows are partly demographic: shrinking working-age populations, Germany's especially, mean workers are scarce there. So the honest reading is that Britain's divergence is substantially a policy story as well as a technology story. But that sharpens the point rather than softening it: the UK has, in effect, subsidised the substitution of AI for entry-level people, and is simply the first place to show what that looks like.
The international evidence points the same way. A Stanford analysis of millions of US payroll records found that, since generative AI spread, employment for early-career workers in the most AI-exposed occupations has fallen by around 13 per cent relative to their peers, even after controlling for firm-level shocks, while older workers in the very same occupations held steady or grew.17 The IMF estimates that around 60 per cent of jobs in advanced economies are exposed to AI.18 And the people building the technology are not reassuring on this point. Anthropic's own chief executive, Dario Amodei, has warned publicly that AI could eliminate half of all entry-level white-collar jobs within one to five years and push unemployment in developed economies to between 10 and 20 per cent.19 That is where the figure in this section comes from, and it deserves a sceptical note of its own: an AI chief executive prophesying his own technology's world-changing power is not a disinterested witness, because capability talk also sells. Treat it not as testimony but as a named scenario from someone with unusual visibility of the technology, one that happens to point the same way as the payroll data above.
Against that backdrop, unemployment of ten or twenty per cent stops being science fiction and becomes the middle of the published range. The IPPR has modelled the "second wave" of AI adoption, the stage at which firms move beyond pilots and embed the technology deep in their processes, and put up to eight million UK jobs at risk in its worst-case scenario, roughly a quarter of the workforce, with entry-level, part-time, and administrative roles most exposed and women and young workers hit hardest.20 If even half of that materialised, two things would happen to therapy at once. The need would surge: unemployment is one of the most reliable predictors of depression, anxiety, and suicide risk that we have, and in the classic meta-analysis unemployed people showed roughly twice the rate of psychological problems of those in work.21 And the means to pay would collapse: private therapy at today's hourly rates is, in practice, a service for the employed. A profession priced for the salaried would find itself facing a population that is anything but.
Notice that these two futures are not alternatives, even though they sound like opposites: one where AI takes therapeutic work away from us, one where the fallout from AI creates more therapeutic need than we could ever meet. They are more likely to arrive together, and they point to the same adaptation: finding ways to help far more people at a far lower cost per person. Supervised AI-delivered care, blended and stepped models, group work, low-cost digital tiers with human oversight. Part of that is an ethical opportunity, because even now most people who need therapy never get anywhere near it; cost, waiting lists, and geography see to that. And part of it, said plainly, is self-preservation. If good-enough AI support exists at a tenth of our price and our clients can no longer afford us, the question of whether AI should do therapy will be answered by their bank balances rather than by our position statements.
Adaptation is not the profession's only lever, and I do not want to write as if market logic were weather. Regulators and professional bodies can set standards for what may call itself therapy, for what AI-delivered care must demonstrate before it touches a distressed person, and for what a registrant's supervision of such care actually requires; registrations that refuse to lend their names to unsafe products are worth more than any position statement. Professions have shaped markets before. But standards shape a market; they rarely stop one, which is why I think we need both: the collective work of setting the bar, and the individual work of being ready.
I hold all of this lightly. Forecasts about AI and employment have a poor track record in both directions, and there are things in the room, the relationship itself, rupture and repair, the experience of being accurately known by another nervous system, that I am not convinced will ever scale. But "I am not convinced it scales" is a hypothesis, not a business plan. The prudent position for a working therapist is to get curious now: learn the tools, understand what they can and cannot hold, and start thinking about what a practice that serves ten times as many people at a fraction of today's price would look like, before the economics forces the question on us.
One scale up: the question my numbers cannot answer
There is a fair objection to everything above, and it deserves its own section rather than a parenthesis: if every individual footprint is negligible, isn't that exactly how every collective problem hides? My practice's arithmetic is the same as every other user's, and several hundred thousand negligible clinicians, plus a few billion negligible everyone-elses, are precisely what the data centres are being built for.
And the aggregate is genuinely large. Gartner forecasts global data-centre electricity consumption of 565 terawatt-hours in 2026, up 26 per cent in a single year, with power demand reaching 132 gigawatts now and an estimated 290 gigawatts by 2030; on their analysis, AI capacity is now constrained by the availability of power itself.22 Goldman Sachs expects US data-centre power demand to double by 2027 and estimates that only around half of the capacity scheduled for the next two years will actually arrive on time, because the grid cannot keep pace.23 The money is on the same scale as the megawatts: the four big hyperscalers alone plan roughly 725 billion dollars of capital spending in 2026,24 and JPMorgan puts the global data-centre and AI infrastructure bill at more than five trillion dollars over five years, financed in growing part by debt.25 Whether that buildout is sustainable, environmentally or financially, is a serious open question, and I will not pretend my fridge-freezer comparisons answer it.
But notice that these are two different questions. "Should one clinician feel guilty about transcribing a session?" is answered by the numbers earlier in this piece, and the answer is no. "Is the industry-wide AI buildout sustainable?" is a question about energy policy, grid investment, and market discipline, and an individual practice's footprint says nothing about it either way. Conflating them does damage in both directions: it paralyses individuals over a rounding error, or it lets the systemic question hide behind reassuring personal maths. They deserve separate answers. Mine are that the practice-level cost is trivial, and that the system-level question is real, open, and belongs to regulators, investors, and energy planners, with citizens, therapists included, pressing them on it.
The bottom line
Set against even a modest, consented clinical benefit, the environmental cost of AI-assisted practice is not a close call. A year of transcribing and summarising five hours a day costs about what a fridge uses in a couple of months and emits less than a short drive. That is comfortably outweighed the first time a client tells you the summary helped them hold onto the work, or the first time you finish a session genuinely present rather than scribbling.
The cost that does require active management is not carbon at all. It is consent, data, and the discipline of keeping the clinician, not the model, responsible for the clinical work. Choose tools on those grounds, get specific and revocable consent, keep generative-in-session AI as decision support rather than autopilot, and let the environmental question take its rightful, modest place in the conversation.
And keep one eye on the horizon. The watt-hours were never the threat; the price signal is. A technology this cheap and this capable will reshape both who can afford therapy and how it is delivered, and the therapists who do well by their clients through that change will be the ones who started adapting before they had to. The planet can spare your watt-hours; the buildout is a question for another scale of decision-maker. The client's trust is the thing to protect, and reaching the many people who need that trust and have never been able to afford it may turn out to be the profession's next job.
Notes on figures. All energy and water figures are inference-only and exclude amortised model training; they are best-available estimates with real uncertainty, and actual costs vary with the models, providers, and settings used. Labour-market figures are as of mid-2026. The IPPR's central scenarios are far smaller than its worst case, which is why this piece treats ten to twenty per cent unemployment as a scenario to prepare for rather than a prediction. Full sources below.
References and further reading
arxiv.org/abs/2404.17394
epoch.ai
cloud.google.com
gov.uk
discoverwater.co.uk
doi:10.1145/3630106.3658996
ons.gov.uk
cityam.com
ajbell.co.uk
ec.europa.eu/eurostat
theguardian.com
peoplemanagement.co.uk
mckinsey.com
lancaster.ac.uk
gov.uk
blogs.lse.ac.uk
digitaleconomy.stanford.edu
imf.org
axios.com
ippr.org
doi:10.1016/j.jvb.2009.01.001
gartner.com
goldmansachs.com
tomshardware.com
datacenterdynamics.com
A note on how this article was written. Fittingly for the subject, it was drafted with AI assistance (Anthropic's Claude), working from my own practice, figures, and views, and reviewed and edited by me throughout. The judgements in it, like the responsibility for them, are mine.