Why Afitpilot Started Giving Sissy Workouts

Over the past month, Afitpilot started producing workouts that were bad.

Not “a bit off.” Bad. Underwhelming, repetitive, low-stimulus sessions that had no business being given to an advanced powerlifter.

Same exercises every week: Romanian deadlifts, face pulls, side planks, ab rollouts. Wrapped in language like “volume base” and “recovery priority.”

Context: 7+ year lifter. Mid-macrocycle. Post-deload. No injury. No fatigue crisis. Zero reason to underload.

Yet the system kept defaulting to safe, generic patterns.

That’s not coaching. That’s avoidance.

People asked: “Isn’t AI supposed to be smarter than this?”

Wrong question. This wasn’t an intelligence failure. It was a decision authority failure. That distinction matters.

---

Where the assumption broke

The exercises themselves weren’t wrong. Romanian deadlifts are fine. Face pulls are fine. They’re reasonable tools.

The problem is context. These are what you give someone in week one, or during active recovery, or when managing an injury. Not an advanced lifter mid-cycle with a green light to push.

So: why was the AI choosing them?

Here’s what I found. When an LLM doesn’t have hard constraints telling it what phase we’re in, whether stress is required, whether underloading counts as failure—it defaults to the lowest-risk option available.

Large language models don’t want progress. They want to not be wrong.

“Safe” in fitness AI looks like: RPE caps, rehab-adjacent movements, shoulder health clichés, core stability clichés, endless base building. General population programming. Defensible to everyone, useful to no one specific.

---

The second failure: no memory

The repetition was the other red flag.

Same movements kept appearing because the system wasn’t tracking exposure history. It didn’t know “you already used this last week.” Every week was a blank slate.

So it kept reaching for the most defensible answers it knew. Over and over.

If you don’t give the model memory, it gives you habits. And those habits always skew conservative.

---

The line I crossed

Here’s the core mistake, stated plainly:

I let the AI decide whether to push, instead of only how to push.

That’s the line.

An AI can fill structure, execute constraints, generate variations, respect boundaries. Fine.

But it should never decide: training phase, risk tolerance, whether progression is appropriate, whether an athlete “needs recovery.”

The moment you outsource that judgment, the output collapses into safety-first mush. The model will always choose plausibility over progress, because plausibility is what it’s optimized for.

---

What’s different now

The AI no longer decides what the plan is. It executes a plan I’ve defined.

Specific changes:

• Mesocycles are hard-locked, not inferred from context
• Deloads are discrete events, not ambient states the AI can invoke
• Minimum intensity floors exist—the AI can’t sandbag
• Exercise exposure is tracked and rotation is enforced
• Repetition without explicit justification gets rejected
• “This could be given to a beginner” is now a failure condition

Here’s what the difference looks like. Same athlete profile, same point in the cycle:

Before (May): Back squat 3×6 @ RPE 7, Romanian deadlift 3×10, face pulls, plank holds. Low demand, no progression signal, cookie-cutter.

After: Back squat 4×4 @ RPE 8 with 3% load increase, pause squats 3×3, barbell rows with tempo prescription, weighted carries. Specific stress, clear intent, appropriate challenge for the training age.

---

Why I’m telling you this

I could have fixed this quietly and said nothing.

But if you’ve been using Afitpilot over the past few weeks and the workouts felt too easy, too repetitive, too generic—you deserve to know why.

This was a real failure. I caught it because I use the product myself and because some of you told me something felt off.

The fix is live now.

If it’s still not right, tell me. That’s how this gets better.