Therapy without a therapist: Why unguided self-help might be a better intervention than guided

By huw @ 2024-04-08T06:01 (+74)

Correction: I identified some errors in my cost model that caused me to slightly overstate the difference between the cost of guided and unguided self-help. I've corrected a lot of this post (you'll see strikethroughs where I have), but not the charts. See the comment for more details.

Summary

1. Guided self-help involves self-learning psychotherapy, and regular, short contact with an advisor (ex. Kaya Guides, AbleTo). Unguided self-help removes the advisor (ex. Headspace, Waking Up, stress relief apps).
2. It's probably 6× (2.5–9×) more cost-effective
   1. It’s about 70% as effective against depression
   2. Beneficiaries use it no less than half as much
   3. It's at least 10–15× cheaper, and might scale sub-linearly
3. Behavioural activation is a great fit for it (That's where you think about things that make you happy and make structured plans to do them)
   1. It’s as effective as CBT and other evidence-based therapies
   2. It’s the strongest significant component of internet-based CBT
   3. It might be easier to self-learn since it's simpler
   4. It might be less stigmatising since it's less medicalised
4. It’s less risky, mostly because it scales better
   1. There’s less evidence overall, but not much
   2. It scales superbly, so it’s highly funding-absorbent
   3. It can fail faster and cheaper
   4. Externalities are small, but displacement is concerning

How much worse is therapy without a therapist?

A lot of work has already been done in EA to emphasise mental health as a cause area. It seems important, tractable, neglected relative to other interventions, and is at least in the conversation for cost-effectiveness^[1]. And unlike many other health issues, we can only expect it to get worse over time^[2].

Guided self-help is an intervention which incorporates self-directed learning of a psychotherapy, and brief, regular contact with a lightly-trained healthcare worker. It can be deployed in highly cost-sensitive environments, and flagship programmes have been developed with the WHO and deployed across Europe, Asia, and Africa^[3][4][5][6]. Off the back of their own research (easily the most comprehensive cost-effectiveness analysis)^[7], Charity Entrepreneurship recently incubated Kaya Guides, who are cost-effectively scaling the same programme in India^[8].

But the same report also notes that removing the guided component might be even more cost-effective. This is called unguided (a.k.a. pure) self-help, and it’s usually defined as any self-learned psychotherapy (regardless of whether that psychotherapy is evidence-based). The early examples involved reading books, such as the Overcoming series, but modern interventions are usually apps, such as Headspace, Waking Up, Clarity CBT Journal, Thought Saver, UpLift or just versions of Step-By-Step and Kaya Guides without the guides. This report’s definition is deliberately broad to keep in line with the cited literature, but when talking about potential interventions I’m generally thinking about apps based on evidence-based techniques.

This report is purely comparative; it's only valuable if you already believe that guided self-help might be a promising, cost-effective intervention. I'll discuss dollar-for-dollar cost-effectiveness, make some arguments for behavioural activation as a uniquely well-suited psychotherapy to apply, and finish up by arguing that since it scales so much better, it's much less risky to try. (Note: That last bit is a bit self-serving since I'm applying to AIM with this idea).

Finally, I'll limit the analysis to depression since it's the most burdensome mental health problem, but many of the included studies also find similar results for anxiety.

Unguided self-help is probably 6× (2.5–9×) more cost-effective than guided

Let's start with cost-effectiveness. Here's my chain of reasoning:

1. It’s about 70% as effective against depression
   1. We should compare against waitlist controls, despite possible bias
   2. We should compare against studies which recruited depressed people
   3. The best meta-analyses show ~70% effectiveness for these conditions
2. The effect shouldn’t decay differently
   1. If therapy effects decay exponentially, we’d see little to no difference between guided and unguided at follow-up
   2. We find no evidence of a difference at 6–12 months follow-up (but there isn’t much data)
3. Beneficiaries use it no less than half as much
   1. The engagement difference is no worse than half in trials
   2. The difference might not be due to expertise
   3. The difference might really be due to encouragement
   4. Real-world unguided engagement isn't obliteratively low
   5. Since real-world conditions are similar, we can extrapolate from trial contexts
4. It's at least 10–15× cheaper, and might scale sub-linearly
   1. In Charity Entrepreneurship's CEAs, removing guidance reduces costs by ~93% at scale
   2. Using Kaya Guides' estimates, removing guidance reduces costs by ~90% at scale
   3. An effective intervention should be able to scale development sub-linearly
   4. Marginal costs might scale to zero due to organic recruitment

In the worst case, the effect size might be around half, with half the engagement, and only 10% of the cost. This would still generate a 2.5× improvement. And in the best case, we might achieve 9× (with 75% engagement and 7% of the cost).

I don't think you need to believe in my exact estimates to agree with me—the lowest bar here would probably be to agree that it's unlikely that the differences between the two delivery forms are an order of magnitude or more, and as we'll see, there isn't strong evidence for this.

It’s about 70% as effective against depression

We should compare against waitlist controls, despite possible bias

86% of active mental health cases in LMICs receive no treatment, compared to 63% in HICs^[9]. The median number of mental health workers per 100,000 population in LMICs is 2, compared with 60 in HICs^[10].

We should expect that almost all recipients of a mental health charity operating anywhere in the world should be untreated. Furthermore, most psychological trials are run in HIC contexts, where treatment-as-usual (TAU) means much more care than in LMICs. This means that we should focus on trials that compare intervention to an inactive control, typically a waitlist. However, there is considerable debate about whether waitlist controls inflate effect sizes^{[11][12][13][14]} (by approx. 25%^[14:1]), so we should still remain mindful of TAU control results.

We should compare against studies which recruited depressed people

More severely depressed people have more to improve, so we shouldn't be surprised when we find that baseline severity moderates effect sizes of internet-based psychotherapies^{[15][16][17][18][19]}. We should be careful not to include studies of general populations, and at least require that the patient score for depression, since a new effective intervention should expect to treat people presenting with a problem, or at least to spend donor money in a targeted way.

The best meta-analyses show ~70% effectiveness for these conditions

A very large network meta-analysis ($k$ = 155, $N$ = 15,191) found unguided self-help has 60% of the effect size of guided ($d$ = 0.52 vs 0.87) for patients with depression against waitlist^[20]. Restricting to just low risk of bias studies ($k$ = 71) improved the gap to 64%; restricting to just internet-delivered self-help ($k$ = 148) improved it to 70% ($d$ = 0.57 vs 0.8), at which point the difference was no longer statistically significant. However, they could not find a significant advantage of unguided self-help against TAU in any condition.

Another very large study of individual data for depressed patients ($k$ = 39, $N$ = 9,751; 8,107 aggregated) found unguided had 76% of the effect size of guided against waitlist ($d$ = 0.43 vs 0.57), and found a significant advantage of unguided against TAU^[16:1]. They only found weak evidence of publication bias, and risk of bias in their included studies was low, so they didn't split out their analyses in the same way.

The effect shouldn’t decay differently

The decay characteristics of psychotherapy are not well understood^[21], and few studies provide long-term follow-up at real-world intervals^[22]. Assuming a null hypothesis of exponential decay, we'd expect the difference to disappear over time (as larger effects will wear off faster). Neither of the two very large meta-analyses above found evidence of a difference between guided and unguided self-help at ≥3 months follow-up^[16:2][20:1], although neither had much data.

Beneficiaries use it no less than half as much

We suspect that there might be a difference between engagement in guided and unguided self-help. In trial contexts, we can measure the amount of content engaged with (adherence), and how many people drop out of treatment (acceptability/dropout/attrition), but extrapolating these to real-world interventions is hard because they don't have upper bounds on engagement or duration. We’ll reason that it’s valid to extrapolate the relative guided-unguided difference to real-world contexts, because the primary factors of that difference hold under both conditions & real-world unguided engagement isn’t terribly low.

The engagement difference is no worse than half in trials

No meta-analyses found a relative risk or odds ratio of more than 2 for the improvement of guided self-help on adherence and acceptability^{[20:2][16:3][23][24]}, and one very large meta-analysis found that guided self-help decreased acceptability^[20:3]. Beyond this, the studies disagreed about the intensity of the association; heterogeneity even in direct comparisons was high ($I^2$ = 42% for adherence and 52% for acceptability^[23:1], where the latter was statistically significant). We can only conclude that the engagement difference should be no worse than half (and likely have little difference), and more driven by adherence than acceptability.

The difference might not be due to expertise

There is no evidence that stronger qualifications improve effectiveness^{[24:1][23:2][25]} or adherence ^[23:3], such that lay mental health workers are highly effective^{[26][3:1][4:1]} and even non-inferior to accredited psychotherapists in large trials^[25:1]. Going further, a very large individual participant data (IPD) meta-analysis ($k$ = 76, $N$ = 17,521; 11,704 aggregated) deconstructed factors relevant to adherence across guided and unguided internet-based self-help^[27]. Neither initial face-to-face contact nor treatment-relevant guidance improved effectiveness or adherence; rather, where human guidance improved those outcomes it was likely due to encouragement and did not depend on expertise.

The difference might really be due to encouragement

As mentioned, the above very large IPD meta-analysis found stronger evidence for both human and automated encouragement on adherence, although neither were quite statistically significant^[27:1]. Automated encouragement also improves acceptability^[28] and effectiveness^[29] (although the latter study only reached significance for anxiety symptoms, not depression). Additionally, making human guidance on-demand reduces adherence and acceptability^[23:4]. Finally, in real-world settings, user encouragement features such as calls to action, adaptivity, and feedback improve 30-day retention^[30], unique sessions^[31], and usage time^[30:1][31:1].

Real-world unguided engagement isn't obliteratively low

Acceptability rates for mental health phone app RCTs average around 74%^[18:1], while real-world 30th-day retention is 3.3% (median; IQR 6.2%; this should heavily underestimate trial acceptability because it only measures usage on a single day, while trials collect follow-up over many weeks)^[32]. In the worst case, when directly comparing trial data of publicly available apps to real-world usage, adherence is only 4 times lower^[33].

Since real-world conditions are similar, we can extrapolate from trial contexts

In trial contexts, we see a solid gap between engagement of unguided and guided self-help. Although precise numbers are hard to narrow down, it's not worse than a 50% reduction. We know that the gap is more likely to be driven by encouragement than expertise, which, if anything, is likely to improve for real-world interventions with more development work; but at least isn't likely to reduce. When we look at real-world engagement, we don't have comparative data for guided and unguided self-help, but we know that unguided self-help's doesn't decrease by an order of magnitude. Therefore, I think a conservative read of the evidence puts the worst case at a 50% reduction in engagement, but with potential for a smaller gap.

It's at least 10–15× cheaper, and might scale sub-linearly

In Charity Entrepreneurship's CEAs, removing guidance reduces costs by ~93% at scale

Charity Entrepreneurship modelled 6 cost-effectiveness analyses (CEAs) assessing different forms of self-help, including one for unguided self-help; the latter outperforming the others in cost-effectiveness^[7:1].

When we remove guidance costs without altering the cost structure or scale, we lower costs by up to 75%.

We must acknowledge that Scenario 6 uses volunteer labour, so treats staff time as free, but real-world Step-By-Step implementers believe this is unlikely to scale^[5:1]. Furthermore, Scenarios 1 and 3 recruit guides from high-income countries, which would inflate costs. Scenarios 3 and 4 use reused workbooks, which accounts for their reduced costs. Scenario 4 is the most representative of those provided.

However, in practice, the cost structure of a Step-By-Step-style programme would is very different, since it doesn't distribute workbooks. We can modify the scenarios as follows:

1. Reduce workbook, delivery, and travel costs to $0
2. Add 1 additional staff member per 250k beneficiaries for app development
   • The approximate number of downloads per employee across top self-guided mental health apps is 85–650k ($\mu$ = 250k; Appendix 1). A self-guided mental health app developer reported reaching 8,000,000 users with 2 developers^[34].
3. Reduce recruitment costs to $0.02–0.10
   • This was reported as $0.02-0.10 by a self-guided mental health app developer as the cost per install for low-income countries^[34:1].
   • Multiply by 2 in the unguided case to account for lower engagement rates.
4. Divide recruitment costs by 10–15% to get the cost per first starter (~12.34% of Kaya Guides installers actually started the programme^[8:1])
5. Break out staff time per participant
   • Each participant has 1 to 10
     • We set the minimum to 1 since the 12.34% of installers above all made it to a call
     • We set the maximum to 10 since a typical programme has about ~8 weeks of calls
   • Staff spend 15 to 40 minutes on each call
     • Some programmes do 15 minute calls^[8:2], others do half-hour calls
     • There's some leeway time per call for notes, chat review, etc
6. Split office space costs to account for different numbers of staff in each condition.

The modified Guesstimate model is in Appendix 2. In it, unguided self-help scales extremely well.

The following plots include the original numbers before I updated the model, but are close enough to the new estimates that it's not worth changing them.

Under this cost structure, the proportion approaches a limiting value of around 7% (credible interval 1.5–22%) at modest scale—a reduction of 93%. At this scale, the costs begin to linearise and beneficiary recruitment becomes the main cost. Even if we assume a fully volunteer-based model is possible, managerial staff & administration dominate guided self-help, leaving the limiting proportion at 14%.

The proportion is sensitive to the number of hours per participant. Reducing the minimum time to 15 minutes (a single Kaya Guides session, excluding buffer time to prepare & review) increases the limiting proportion to 8.5%. However, increasing the maximum to 780 (a 15-minute session per week for a year, excluding buffer time) a session per week per year reduces it to 4.8%. The issue here is that decreasing staff hours per participant is likely to proportionally reduce effectiveness, to the point of realising that the most cost-effective intervention eliminates guidance entirely.

Using Kaya Guides' estimates, removing guidance reduces costs by ~90% at scale

At 15,000-person scale, Kaya Guides estimate a total cost of $9.32 per beneficiary, and a marginal cost of $3.93 for new beneficiaries^[8:3]. The modified Charity Entrepreneurship model is close to agreement, predicting a total cost of $13–35 per beneficiary and a marginal cost of $3.40–15 per beneficiary. If we divide the marginal cost of treating a guided participant by 2, to agree with Kaya Guides, it predicts a cost reduction of 44% for unguided self-help.

However, this occurs at Kaya Guides' predicted 3-year scale. At their predicted 5-year scale of 100,000 beneficiaries, the reduction increases to 78%, and limits at around 90%.

An effective intervention should be able to scale development costs sub-linearly

It's not clear how to account for development costs. Although major software companies tend to scale their headcount linearly with the number of users, an unguided self-help intervention only needs to be developed once and delivered to multiple contexts, only incurring adaptation costs for new target populations (as Step-By-Step did^{[35][6:1][36]}). Once an intervention reliably produces an effect size, additional development isn't necessary and a smaller group of engineers could handle maintenance and support, which tends to scale sub-linearly as bugs are fixed without new ones being introduced.

Marginal costs might scale to zero due to organic recruitment

As the intervention proliferates, it will appear higher in searches and app store rankings, and spread through media and word of mouth. As such, at higher scale the cost of recruiting new beneficiaries should decrease—and these beneficiaries should be less likely to drop out. A self-guided mental health app developer reported an organic install rate of roughly 2/3rds, and speculated that popular apps would reach close to 100%^[37]. Practically, the marginal cost of recruitment may even approach zero.

Behavioural activation is a great fit for unguided self-help

Cognitive behavioural therapy is made up of many components, and although the overall programme is highly effective^[38][22:1], component analyses have struggled to determine what drives it^{[39][40][41][27:2]}. We can use this to our advantage by identifying components that are likely to be non-inferior, but better suited to unguided self-help.

One such component is behavioural activation therapy (BAT; itself a subset of CBT). It is extremely simple, as it involves tracking one's mood, identifying pleasant activities and creating structured plans to do them^{[42][43][44][41:1][45]}. Although developed by Lewinsohn in 1975^[44:1], BAT fell out of favour during the cognitive revolution before being revived in 1996 by a study demonstrating its non-inferiority to pure cognitive therapy and cognitive behavioural therapy^[41:2].

In this section, I'll argue that behavioural activation:

1. Is as effective as CBT and other evidence-based therapies
2. Is the strongest significant component of internet-based CBT
3. Might be easier to self-learn since it's simpler
4. Might be less stigmatising since it's less medicalised

It's as effective as CBT and other evidence-based therapies

In the most comprehensive meta-analyses, BAT is highly effective against active and inactive controls ($k$ = 331, $N$ = 34,285^[38:1]; $k$ = 26, $N$ = 1,524^[46]; $k$ = 16, $N$ = 780^[45:1]) and non-directive counselling^[38:2], non-inferior against all other evidence-based psychotherapies in effectiveness, acceptability, and remission rates^[45:2][38:3], and specifically non-inferior against pure cognitive therapy and cognitive behavioural therapy^[45:3] ($k$ = 45, $N$ = 3,382^[47]).

It's the strongest significant component of internet-based CBT

A large meta-analysis of component studies found removing behavioural activation from CBT significantly reduced effect sizes, although only had 3 direct comparisons^[39:1]. Another large network meta-analysis ($k$ = 76, $N$ = 6,973) found a non-significant moderate depressive effect of behavioural activation, but it only had a few arms without it, and only measured face-to-face CBT^[40:1].

However, a very large IPD meta-analysis ($k$ = 76, $N$ = 17,521; 11,704 aggregated) found a significant positive effect of behavioural activation using 84 out of 179 arms ($d$ = 0.13–0.48), making it the strongest analysed component^[27:3]. As well as having more contrasting arms, by using internet-based CBT it didn't risk face-to-face therapists incorporating elements from other therapies ad-hoc, so was more statistically powerful. The authors note that this restriction hampers extrapolation, but for unguided self-help it's clearly superior to other therapies.

It might be easier to self-learn since it's simpler

As discussed, there isn't strong evidence that qualifications improve effect sizes^{[24:2][23:5][25:2]}. Further, a study directly compared CBT administered by a licensed psychotherapist with BAT administered by lay mental health workers, and found no difference between treatment arms ($N$ = 440)^[25:3]. The very large IPD meta-analysis above concluded that behavioural activation is more effective in self-help contexts specifically because it's easier to self-learn^[27:4], and this concurs with intuitions about its simplicity.

It might be less stigmatising since it's less medicalised

Mental health stigma is higher in low-income countries^[48][49], which drives self-stigma^[50][51][52] and reduces mental health care utilisation^{[51:1][53][54]}. High assessments of public stigma are associated with higher dropout in guided and unguided self-help, although not adherence^[55].

Unguided self-help has been shown be an acceptable intervention for people otherwise unwilling to engage with face-to-face treatment^[56]. When Step-By-Step was trialled in Lebanon, nearly half of the participants didn't tell anyone they knew for fear of stigmatisation^[5:2], or even partner violence^[6:2] (if anything, however, this indicates that it was an acceptable & private intervention for many). One Syrian refugee noted:

When you hear the word psychiatrist, it is associated with craziness or something.^[6:3]

We can reason that behavioural activation may be less stigmatising over and above unguided self-help. Unlike many other therapies, it can be delivered without psychoeducation or even requiring the recipient to admit they have a mental health problem. The intervention can be designed to minimise medicalised terms, which Step-By-Step did in Lebanon^[6:4] and was suggested by Step-By-Step participants in Germany and Sweden as needing further improvement^[57]. In lower-income countries, users mentioned sharing phones and wanting privacy^{[6:5][36:1][58]}; the plausible deniability of a less medicalised app may help with this. However, placebo effects may be a large part of psychotherapy's effectiveness^[27:5][38:4], so demedicalisation may not be worth the trade-off.

It’s less risky, mostly because it scales better

For the sake of convenience, this section will mirror the Charity Entrepreneurship report's risk factors: idea strength, limiting factors, execution difficulty, and externalities. I’ll argue:

1. There’s less evidence overall, but not much
   1. Most of the evidence emerged since CE recommended guided
   2. There’s now an effective BAT app intervention that just needs to be scaled
2. It scales superbly, so it’s highly funding-absorbent
3. It can fail faster and cheaper
4. Externalities are small, but displacement is concerning

Overall, the value of higher scale—both directly, and as a de-risker—should outweigh the small differences in the other factors, which would make unguided self-help a better intervention.

There's less evidence overall, but not much

Although the literature generally produces statistically significant results for unguided self-help, face-to-face therapy will always have stronger support. Putting aside all of the evidence, it's hard to shake the feeling that this shouldn't work as well as it does, that perhaps unguided self-help completely collapses outside a trial context. This is not an insignificant risk, and I don't wish to downplay it.

Most of the evidence emerged since CE recommended guided

Most of the evidence base for unguided self-help and behavioural activation is very recent, and has emerged since CE’s report recommended guided self-help^[59]. The largest meta-analysis of mental health apps was repeated with the same criteria in 2017, 2019, and 2024, finding 18, 66, and 176 RCTs respectively^{[60][29:1][18:2]}. As such, the most downloaded evidence-based mental health apps are almost exclusively based around mindfulness meditation, with a handful of mood trackers, CBT psychoeducation, and relaxation apps^[32:1], because the evidence for better-suited techniques & components didn’t exist. It should be possible to surpass these in effectiveness by incorporating new evidence.

There’s now an effective BAT app intervention that just needs to be scaled

Many of the most promising interventions focus on scaling something that has been directly successful in clinical trials. Kaya Guides, for example, is just adapting & operationalising Step-By-Step^[8:4], which has already proven to be successful^[3:2][4:2]. A big de-risker is that such a treatment already exists; the Moodivate app is an unguided behavioural activation intervention and is effective^[61][62][63], adhered to^[61:1][63:1], and culturally adaptable without losing these properties^[63:2]. All of the discussed individually effective components appear to come together in practice, although the evidence base is still much more limited than for equivalent guided- and CBT-based interventions.

It scales superbly, so it’s highly funding-absorbent

Software has scaled to envelop the world faster than any innovation before it. New apps can reach hundreds of millions of users within days of launching^[64]. This is because software, by and large, doesn't require any marginal cost to reach new users (although businesses may still spend on customer acquisition). They are targets for huge venture capital investments precisely because their upside is nearly uncapped.

Guided self-help, however, has a strong limiting factor in guidance. Guides not only have to be recruited and trained, but need ever-deepening layers of management. All of this is likely to heavily lower funding absorbency, since hiring & scaling in anticipation of future growth isn't cost-effective, but doing it reactively is slow. To illustrate, Kaya Guides estimate a decreasing growth rate over time, but in a best-case scenario where they continue to double annually after year 5^[8:5], it will take 15 years to reach 100 million people (unguided could get there 1,000 times faster). An unguided intervention, on the other hand, could theoretically get there 1,000 times faster, because its only practical limiting factor is ad spend for recruitment.

Not only should cost improve at scale, but effect size & engagement should too. Advanced software apps are able to use analytics across huge audiences to automate experimentation against both of these metrics. One expert app developer believed the best apps may be 1–2 orders of magnitude more effective^[34:2], concurring with wide between-app variance in trials^[18:3], and outlier engagement numbers for the largest mental health apps^[65].

It can fail faster & cheaper

The corollary of faster scale and lower costs is that it can fail faster, and for cheaper. The structure of an unguided behavioural activation app is particularly advantageous, as beneficiaries are recruited via ads and log their mood daily. The impact of changes on effect size & engagement can be measured within days, and a preliminary trial with those outcomes can be produced faster and with higher scale than the years it takes typical new charities. As such, the major path to failure is that a Moodivate-style app fails to reproduce a strong effect size with a real-world cohort (rather than running into logistical or policy issues).

Since it’s free to expand an app to high-income countries, if an intervention fails to achieve cost-effectiveness quickly it could self-fund indefinitely. Even if it shuts down, it could contribute private aggregates of the data to the scientific community^[66], in what, at even modest scales, would be the largest-ever study of global mental health & wellbeing.

Externalities are low, but displacement is concerning

Unguided self-help is less likely to be harmful than control conditions. Both guided^[67] and unguided^[68] self-help have been shown to reduce deterioration rates against inactive controls, although the evidence for the effect of unguided against TAU was insignificant^[68:1]. Neither study found a subgroup or moderator for whom self-help would be worse than inactive controls; it does seem to be better than nothing.

The bigger risk is that unguided self-help displaces more effective treatments, both for individual recipients, and for funding. The corollary of unguided self-help having lower stigma is that some recipients may prefer it over face-to-face treatment, even when the latter would be better & more cost-effective for them.

If an unguided intervention demonstrated a higher cost-effectiveness, it might draw funding away from more effective charities such as Kaya Guides, and from the hard governmental work of building out mental health infrastructure. This is probably bad. Hopefully in an Effective Altruist context, pushing the state of the art on cost-effectiveness would draw enough new funding into the space to offset the displacement impact.

Conclusion

New technology may be able to close the effectiveness and engagement gaps further. One meta-analysis found weak evidence that rudimentary chatbots improve effectiveness of mental health apps^[18:4]. Where there are obvious legal concerns, modern AI techniques still have the potential to dramatically improve personalisation without being embedded in a chat UI, or to culturally & linguistically adapt interventions to new contexts. Being ready for cloud and on-device costs to come down over the next half-decade should be a key consideration of any unguided intervention.

I'm less sure about what form such an intervention would take. It might make equal sense as a non-profit, or as a for-profit that donates a large proportion of its profits to subsidising itself in LMICs. The latter is particularly tempting, because it could be a target for venture capital, which brings money and decades of expertise in scaling software products—but this would equally bring the risk of trying to juice profits from countries that need such an intervention to be free.

To finish up, I'll spare the pretence: I think it's really unintuitive that you can do therapy without a therapist. What keeps me confident is that everything I've discussed is relative to guided self-help (an already established intervention) and the credible intervals for unguided self-help are well above guided's. If you believe guided self-help is likely to be very effective, then I hope it's not a big step to believe the same about unguided, and moreover, that there's barely any cost to finding out.