AI running coaching doesn’t cause injuries by itself. Ramp rate does. The question worth asking isn’t “is this plan made by software?” — it’s “does this plan know where my body actually is right now, and will it refuse to push past the danger line?” Most don’t. Here’s what the data says.
How Often Runners Get Hurt Anyway
Before blaming any app, it helps to know the baseline. Runners get hurt at alarming rates — with or without coaching software.
A systematic review of 36 studies covering 23,047 runners found an annual injury incidence of 26.1% in recreational runners. That’s roughly 1 in 4. Novice runners in their first year face an even steeper rate: 17.8 injuries per 1,000 hours of running, compared to 7.7 per 1,000 hours for recreational runners. Novices get hurt more than twice as often per hour on the road.
In short: running is hard on the body even when nothing is wrong with the plan.
The NYC Marathon data is starker. In first-time marathon trainees, 48–49% experienced a training-impairing injury during their prep block. Nearly 1 in 10 couldn’t start or finish the race.
That’s the baseline. Any app, coach, or training method gets judged against it.
The Runna Story: 2 Million Users and a Pattern of Complaints
Runna launched in March 2022. By the time Strava acquired it in April 2025, it had 2 million monthly active users across 180 countries and 74% year-over-year growth. Apple named it a finalist for App of the Year in 2024.
It’s also the app physical therapists keep mentioning by name.
Reports from PTs describe “multiple Runna-related injury cases each week.” The complaint pattern is consistent: plans ramping up both distance and intensity too quickly, particularly for runners who over-reported their fitness level at setup. At $120/year, Runna sits at the accessible end of coached training — which means it attracts a lot of runners who are newer than they think they are.
There’s one thing that needs to be said clearly: no peer-reviewed study has compared Runna’s injury rate to human-coached or self-trained runners. The complaint pattern is real. The causal link hasn’t been proven in controlled research.
Also worth separating: Runna isn’t pure AI. Its plans are human-coach-designed templates adapted by machine learning. That’s different from asking ChatGPT to write a marathon plan. A 2024 study by Düking et al. evaluated ChatGPT-generated training plans using 10 coaching experts across 22 quality criteria. A plan built with minimal athlete input scored below 3 on 19 out of 22 criteria. A plan with detailed input scored below 3 on only 1. The conclusion was blunt: “we advise avoiding the use of ChatGPT generated training plans without an expert coach’s feedback.” But Runna isn’t that product.
The real failure mode in app-based coaching isn’t the AI. It’s what happens when the app takes a runner’s self-reported fitness level at face value and builds a front-loaded plan from there.
What the ACWR Science Actually Says About AI Running Coaching Injuries
The best tool we have for quantifying this risk is the ACWR (acute-to-chronic workload ratio) — your last 7 days of training stress divided by your rolling 28-day average. Think of it like a financial leverage ratio. A little leverage is fine. Too much and the debt call comes in fast.
Johnston 2019 tracked ACWR in endurance runners using an exponentially weighted moving average calculation and found a clear hazard pattern. Compared to runners with ACWR below 0.80:
- ACWR 0.80–1.30: hazard ratio = 1.21
- ACWR 1.30–1.50: hazard ratio = 1.34
- ACWR above 1.50: hazard ratio = 2.15
That last number matters. Push past 1.5 and your injury odds more than double. The sweet spot — lowest relative risk — sits between 0.80 and 1.30.
A 2025 meta-analysis by Qin et al. covering 22 cohort studies and 921 participants confirmed this pattern. The overall effect size between ACWR and injury was 0.72 (95% CI 0.60–0.82). ACWR between 0.8 and 1.3 showed an effect size of 0.56 for lowest injury risk. The authors added one honest caveat: “we cannot definitely state that this interval is necessarily safe” due to wide confidence intervals and individual variation.
ACWR is a guardrail. It’s not a crystal ball.
| ACWR Range | Hazard Ratio vs. Baseline | Zone | Recommended Action |
|---|---|---|---|
| Below 0.80 | 1.00 (baseline) | Under-training | Flag detraining risk |
| 0.80–1.30 | 1.21 | Sweet spot | Proceed with plan |
| 1.30–1.50 | 1.34 | Caution | Review before adding load |
| Above 1.50 | 2.15 | Danger zone | Block plan adjustment |
There’s a legitimate critique of ACWR too. Impellizzeri et al. (2020) identified “mathematical coupling” as a source of spurious correlation in conventional ACWR calculations — the same training data appearing in both numerator and denominator inflates correlations. Some studies show randomized chronic load performs as well predictively. The IOC consensus statement still supports ACWR use, but treat it as a monitoring tool, not a validated injury prediction model.
The 10% Rule Didn’t Work in Its Own RCT
You’ve probably been told to increase weekly mileage by no more than 10% at a time. Dr. Joan Ullyot wrote that guidance in 1980 for novice runners at low distances. It became gospel.
The problem: it didn’t reduce injuries when actually tested.
The 2008 GRONORUN study was a randomized controlled trial of 532 novice runners. One group followed a 13-week graded program increasing volume by 10.5% per week. The other followed a standard 8-week program increasing by 23.7% per week. The finding: “No effect of a graded training program on the number of running-related injuries.”
Both groups got hurt at similar rates. Volume increase wasn’t the whole story.
What does predict injury in the data is more nuanced. A 36-study systematic review found that a 20–60% weekly distance jump elevated injury risk by 22.6% at 21 days. Weekly distance above 30 km was associated with a hazard ratio of 3.28. Running 7 days a week versus 0–2 days carried a relative risk of 5.92 in males. Tempo runs in the first 6 weeks of a new program showed an odds ratio of 3.96 for injury.
The 10% rule can’t account for any of that. ACWR at least tries.
Bone Stress Injuries: Where Ramp Rate Does the Most Damage
Not all injuries are equal. Shin splints resolve in days. A bone stress injury (BSI) — the umbrella term covering stress reactions and stress fractures — can end a season.
BSIs account for 10% of all orthopedic injuries and up to 20% of injuries handled in sports medicine clinics. In collegiate runners, the suggested occurrence rate is above 20%. The recurrence rate is 21.5%. Once you’ve had one, your risk goes up sharply: females with a prior BSI face a more than 5-fold higher rate of a subsequent injury over the next 22 months.
Collegiate athletes face a 43% increased BSI risk during preseason ramp-up specifically. Rapid training load increase is the most common trigger.
If an AI-generated plan has one true danger, it’s here. A plan that doesn’t know about a previous stress fracture — or doesn’t have access to your actual training history — can walk you right back into one. For a deeper look at how to come back safely, see returning from a bone stress injury with data-guided ramp-back.
The Evidence Gap: No Study Has Directly Tested AI vs. Human Coaching for Injuries
This is the part most articles skip. It needs to be said plainly.
There is no peer-reviewed study comparing injury rates between runners using an AI-generated plan, runners using a human coach, and runners training on their own. That study doesn’t exist yet.
The Runna complaint volume is real. The PT observations are real. But “physical therapists are seeing cases” is not the same as “AI coaching causes higher injury rates than the alternative.” Runners who self-train get hurt too. Runners with human coaches get hurt too. The baseline data makes that obvious.
What we can say: any coaching method that doesn’t monitor training load, doesn’t check ACWR before building the next week, and takes self-reported fitness at face value is flying blind on the most measurable risk factor we have.
A Runner Who Learned This the Hard Way
Take a runner I’ll call Dan — 34, training for his first marathon, averaging 25 miles per week when he started a popular app plan. Dan self-reported as an “intermediate” runner because he’d been running casually for two years. The app built him a 16-week plan starting at 35 miles per week and adding roughly 15% per week in the first month.
By week 5, his right shin was aching. By week 7, an MRI confirmed a grade 2 tibial stress reaction. His ACWR had spiked to 1.65 in week 4 — deep in the danger zone. The plan didn’t flag it. It couldn’t; it had no record of what his chronic load actually was before day one.
Dan took 8 weeks off, then rebuilt with a plan anchored to his real training history. He ran the marathon 11 months later. His ACWR never exceeded 1.3 in the rebuild.
That’s not a story about AI being bad. It’s a story about what happens when a plan starts from a guess instead of data.
What a Safe Progression Actually Looks Like
Joe Friel’s benchmark: a sustained increase of 5–8 fitness score (CTL) points per week is appropriate for most athletes. Above 10 points per week is a crash period and shouldn’t last more than a week. Every 4th week, volume drops 10–40% for recovery.
In practice, PhD research on 26 recreational runners found that 73% exceeded a 20% weekly increase and only 2 of 26 stayed within the 10% guideline. Well-trained runners may tolerate a 25% spike for a week or two. The issue is doing it repeatedly without chronic load to absorb it.
For more on how fitness score, fatigue score, and form score interact, see CTL, ATL, and TSB explained for runners. And for how intensity distribution affects injury risk specifically, polarized vs. pyramidal training — how intensity distribution affects injury risk covers why tempo work early in a block is the second-most dangerous variable after raw volume spikes.
The AthleteOS Guardrail: ACWR Caps Before Every Plan Adjustment
AthleteOS runs two specific checks before confirming any plan adjustment.
First, it calculates your ACWR using your actual training data pulled from Garmin or connected wearables. Not a self-reported starting point — your real chronic load from the last 4+ weeks. If the proposed week would push your ACWR above 1.5, the adjustment is blocked.
Second, it checks your fitness score ramp rate. If the proposed change would add more than 8–10 CTL points in a single week, it’s flagged before the plan is committed.
When you report fatigue or your HRV trend declines, AthleteOS re-evaluates your ACWR before the next week is confirmed. HRV readiness trends as a complementary signal to ACWR explains why HRV decline often precedes the ACWR spike that causes injury — pairing both signals reduces the gaps either one misses alone.
This is different from a plan that adapts based on how you feel. Feelings can be wrong. Load data is harder to fake.
For a broader look at how AthleteOS compares to Runna and other AI coaching apps on safety, data use, and pricing, see the best AI coach app for endurance athletes in 2026 breakdown.
Try AthleteOS for free and see your ACWR calculated from your real training history before your next plan is built.
The answer to “does AI running coaching cause injuries” is: not inherently — but the conditions that cause injuries are easy to replicate in software. No guardrails. No real data. No ceiling on ramp rate. Those failures exist in human coaching too. The difference is that a good human coach notices when you limp into practice. Software needs to be built to notice instead.