What AI gets wrong about fitness coaching

2026-05-23 · 6 min read

AI is genuinely useful for some parts of training. It's also being oversold for parts it can't do. Here's the honest list of what AI can't replace — and why that's not a problem to solve.

The fitness industry has spent two years promising that AI is about to replace personal trainers. The actual progress has been more modest and more interesting: AI is genuinely good at some parts of coaching and genuinely terrible at others, and confusing the two has cost real money for users who bought "AI coach" apps that promised the wrong things.

At TrainForge we use AI for the parts where it actually helps and explicitly skip the parts where it falls apart. Here's the list of what AI cannot do well — and why pretending it can is harmful, not just inaccurate.

1. Watching your form on a phone camera

The pitch is irresistible: prop your phone against a water bottle, do a set of squats, get instant form feedback. Several apps promise exactly this. The technology — pose estimation from 2D video — has improved a lot. It's still nowhere near good enough to coach a working set.

Camera angle dramatically changes what the model "sees." A 30-degree shift in phone position can flip a perfect rep into a "knee valgus warning."
Lighting, clothing, and background clutter degrade detection. A phone in a busy commercial gym produces noisy data.
The model can't see what matters most for lifting: bar path, joint torque, breathing, intra-abdominal pressure, the look on your face when something is about to fail.
False positives are dangerous. A bad cue ("push your knees out more") on a lift where your knees are fine can cause injury. A coach who says "you're good, keep lifting" rather than risk bad advice is better than an algorithm willing to guess.

TrainForge skips this entirely until the tech is good enough to be safe. We'd rather have no form feedback than wrong form feedback. The bar for replacing a coach's eyes is high, and the tech isn't there yet.

2. The "freeform AI chat coach"

A chatbot you can text any fitness question 24/7 sounds like the ultimate coaching upgrade. In practice, freeform AI coaching produces three predictable failure modes:

Confidently wrong medical advice. "I have lower back pain when I deadlift" might get a response like "try Romanian deadlifts instead" — fine for some causes of back pain, dangerously wrong for others. AI can't differentiate a stiff hip flexor from a herniated disc from a kidney issue, but it'll happily answer all three the same way.
Sycophantic over-validation. Language models are trained to be agreeable. Ask "should I add 30 lbs to my squat this week?" and you'll often get encouragement rather than the honest "absolutely not." A good coach pushes back; a bad chatbot doesn't.
Lost context across conversations. The chatbot doesn't remember that you have a chronic rotator cuff issue from a session three weeks ago unless you mention it again. A human coach holds that context permanently.

TrainForge's coaching uses structured prompts with explicit safety constraints, not freeform chat. The trade-off: it can't answer every random fitness question, but it can't accidentally hurt you either.

3. Knowing when life is the reason

A human coach who's worked with you for six months knows that when you skip three sessions, it's usually not laziness — it's the new baby, the work deadline, the depression flare, the parent in the hospital. The coaching response is completely different for "you're being lazy" versus "your life is hard right now." AI can see the missed sessions, but it can't see the why.

This sounds like a small thing. It's not. The single largest predictor of long-term training success isn't program quality, equipment access, or nutrition — it's whether someone notices when you're struggling and shows up for you anyway. AI can ping you with a notification. It can't be a person who actually knows you.

4. Reading the room on competing priorities

"I want to gain muscle, lose fat, run a half-marathon in 8 weeks, and feel less tired" is a set of incompatible goals. A coach negotiates: pick two; here's the order; let's revisit in a month. AI tends to either pick one goal arbitrarily or produce a milquetoast plan that hedges across all of them, neither of which is what you actually need.

Telling someone "no" is one of the most valuable things a coach does. AI is bad at "no" because it's trained to be helpful. Asking a chatbot whether your unrealistic plan is unrealistic usually gets you a "great question, here are some things to consider" instead of a flat "no, drop one of those."

5. Spotting the lift you're afraid of

A handful of lifts (heavy back squats, heavy bench, heavy overhead, certain deadlift variations) are meaningfully safer with a competent spotter — both physically (catching a failed rep) and psychologically (the difference between attempting a true 1RM with backup versus alone). AI can't spot you. The recommendation to "do heavy singles" assumes you have someone trustworthy nearby, which not everyone does.

TrainForge's program generation accounts for this in two ways: it avoids programming high-risk solo loading (heavy benching without a spotter), and it offers alternative exercises when the safer option costs little. But the underlying limit — that some training works best with a human present — doesn't go away.

6. Holding you accountable in a way that actually works

Accountability is the most overhyped and underdelivered feature in fitness apps. The honest version: nobody actually changes behavior because their phone sent them a streak-protection notification. People change behavior because someone they care about is going to ask them how it went, and they don't want to disappoint that person.

AI accountability is fundamentally fake because the AI doesn't care if you succeed. You know this. Your subconscious knows this. The motivational push notification has the same emotional weight as a billing reminder. A trainer, a workout partner, a coach who's invested in your outcome — those create real accountability because the relationship is real.

What AI is genuinely good at (the honest list)

To balance the criticism: AI in fitness is real value, just narrower than the marketing claims. The honest list:

Pattern recognition in large amounts of data. Did your bench progression slow when you cut your protein? AI catches that. Most people miss it.
Removing manual data entry friction. Photo your meal, get macros. Generate a program in 30 seconds instead of an hour. These save real time.
Personalization at scale. Building a custom program for every member of a gym is impossible by hand. With AI, it's the default.
Consistent application of evidence-based principles. AI doesn't get bored, doesn't have favorite exercises, doesn't drift toward bro science. It applies the same rigor to everyone.
Real-time numerical adjustments. "Your weight is trending down 0.5 lb/wk faster than your goal — drop the deficit by 100 calories." A human coach could do this, but won't every week. AI does.

The right framing

AI in fitness isn't going to replace coaches. It's going to make the coach-plus-tool combination better than either alone. The coach handles the relationship, the context, the form, the spotting, the accountability, the hard conversations. The AI handles the data work, the patterns, the personalization, and the parts that benefit from being relentlessly consistent.

TrainForge is built around that division of labor. If you have a coach, our tools should make them more effective. If you don't, the tools should fill the gaps they're actually good at filling — and we should be honest about the gaps they can't.