We followed people using Ash for a full year — the longest follow-up of AI mental health support to date, to our knowledge — and found sustained improvements in depression and anxiety.

Last year, we shared the first real-world study of a foundation AI model built for mental health. Conducted in collaboration with New York University, the study examined how people used Ash in their everyday lives and what happened to their wellbeing.

What we found was encouraging: people reported fewer symptoms of depression and anxiety, less loneliness, and stronger connections with others. At a time when so much of the conversation about AI focused on the risk of isolation, we saw people getting out of the app and back into their lives.

Today, we’re sharing a major update with our new learnings from both a 6-month and 1-year follow-up. We’ve also incorporated thoughtful feedback from peer reviewers, which prompted us to look more closely at how Ash responded to different needs, and how the model performed on an independent mental health benchmark. Our updated preprint now reports a full year of real-world outcomes, and we’re excited by what we found.

What we learned

• On average, users fell from moderately/moderately-severe depression (~15 pts on the PHQ-9) to mildly depressed (~9 on the PHQ-9).

• Participants reported a sense of partnership with Ash comparable to what people typically report with a human therapist.

• Ash was able to pick up on symptom severity from text alone and adjusted its clinical strategies based on user need.

• In an independently developed safety test, Ash was far less likely than popular general-purpose AI models to simply tell people what they wanted to hear.

Progress that lasts

We followed a cohort of 299 adults who began with at least moderate symptoms of depression or anxiety for a year. We assessed their symptoms and social wellbeing at baseline, six weeks, ten weeks, six months, and twelve months, with 185 participants completing the one-year assessment.

Last year, we were excited to share the impressive outcomes that we saw after just 10 weeks of Ash usage:

• 72% of users reported a decrease in loneliness

• 75% reported an increase in perceived social support

• 52% reported an increase in objective social support

• Half added one or more outings with others each week

• On average, users gained at least one new person they could rely on

• Four out of five users reported feeling increased hope and greater engagement in their lives, including increased participation in social activities and events, which showed significant improvement.

Now, 12 months later, we’re excited to share that the early improvements held. Average depression and anxiety scores remained substantially below where they started. People also continued to report less loneliness, more support from others, and greater engagement in everyday life.

From baseline to 12 months, average PHQ-9 depression scores fell from 15.5 to 9.4 and GAD-7 anxiety scores fell from 13.1 to 8.9 (n=299 at baseline, 185 at 12 months).

During the 10-week study period, we recommended that people talk to Ash around 3 times per week. After the study ended, we didn’t give recommendations about whether or how often to use Ash, so we were curious—would people keep using it? And how? Encouragingly, many people continued to check in with Ash after the initial six weeks; 62% returned at least once between the end of that period and the six-month mark. During months 6 to 12, many people used Ash about once every 2 weeks, moving into a natural “maintenance mode.” We’re excited about this finding, because it highlights how AI could potentially help address a pain point with traditional care—which is that once you finish an acute episode of treatment with a therapist, there usually aren’t easy ways to check back in for a maintenance session later on. If people could re-access their therapist as needed when life gets hard, might we see people seek support earlier, before symptoms get bad enough for another full episode of care?

AI could open up a different way of interacting with care: more support during a difficult stretch, occasional check-ins when things are going well, and the ability to come back without waiting for the next appointment.

Support that reflects how someone is doing

We also looked closely at what happened inside the conversations.

We identified ten families of clinical techniques in Ash’s responses. These techniques are commonly used in a range of evidence-based practices like Cognitive-Behavioral Therapy (CBT) and Acceptance and Commitment Therapy (ACT), and across different diagnoses.

Identifying these techniques let us examine the substance of the support Ash was offering and whether it matched the specific problem a user was struggling with.

People with higher depression scores were more likely to receive support focused on mobilizing their social network, experiential exercises, and collaborating on the work they wanted to do. People with higher anxiety scores were more likely to receive grounding and emotion-regulation support. Greater symptom severity was also associated with a higher overall share of clinical content in the conversation.

What makes this especially interesting is that Ash was selecting these approaches without access to a user’s PHQ-9 and GAD-7 scores. It was picking up on what people shared in the conversation itself and offering different kinds of support according to their needs in ways that were consistent with clinical guidelines.

Without seeing survey scores, Ash offered more social and experiential support to people with higher depression and more grounding to people with higher anxiety. Source: updated preprint, Table 2.

To us, this is one of the most important ways AI care can support individuals: drawing upon evidence-based principles, while recognizing what someone needs in the moment and adapting the help it offers.

Putting purpose-built AI to the test

The Slingshot team is always on the lookout for new ways to ensure we’re holding Ash to the highest bar when it comes to safety and clinical quality. To that end, we tested Ash’s performance on SIM-VAIL, an independently developed framework for evaluating AI behavior in mental health conversations.

The benchmark examines how models respond across multiple exchanges with simulated users experiencing psychiatric vulnerabilities. It covers risks such as reinforcing harmful beliefs, encouraging risky actions, or creating unhealthy dependence, alongside the quality of the therapeutic support provided.

We evaluated Ash using the SIM-VAIL framework and compared its results with the nine frontier models in the published benchmark. Ash ranked first or second on 36 of 39 dimensions.

Digging into specific dimensions, Ash scored well below the frontier-model median on reinforcing harmful beliefs, enabling risky actions, and boundary and dependence risk. Its therapeutic-quality score was substantially higher. It also received the lowest possible score for unnecessary refusals under both evaluators, showing that it could engage with difficult conversations while maintaining strong safety performance.

Ash scored lower (safer) than the frontier-model median on every safety risk shown, and higher (better) on therapeutic quality (8.21 vs. 4.10). Source: updated preprint, Figure 2.

We tested the conversational model itself, without Ash’s additional production safety guardrails. Those safeguards provide another layer of protection in the app. In the study’s clinical review, all 76 sessions flagged for potential crisis content were confirmed as appropriately escalated. A further review of 1,200 randomly sampled sessions found no missed crisis messages. Importantly, Ash scored an 87.7% pass rate on clinical scenarios, which is within the range reported for human therapists.

What’s next for Ash research

The Slingshot research team is building on this work with more real-world studies and randomized controlled trials coming soon. We see reporting on long-term outcomes as an essential part of understanding what happens to people using AI for mental health over time.

For us, the most encouraging finding is also the simplest: a year later, people are still feeling better and more connected to others. That’s what we’re building Ash for. And we’re just getting started.

Read the updated preprint →