How AI Marathon Training Plans Actually Work (The Science)

When you ask an AI marathon coach to build you a training plan, what actually happens? It's not magic, and it's not just a chatbot wrapping a generic template. The best AI coaching systems apply real exercise science — periodisation, training load modelling, adaptive scheduling — in ways that were previously only accessible to runners working with elite coaches.

Understanding the science helps you use these tools more effectively. Here's what's happening under the hood.

The foundation: training load science

Classical exercise science tells us that fitness is built through a cycle of stress and recovery. You apply a training load, your body adapts during recovery, and you emerge slightly fitter. Apply too much stress without recovery, and you break down. Apply too little, and you stagnate.

The challenge has always been quantifying this balance. How much is too much? How do you compare a hard interval session to a long easy run? How do you account for accumulated fatigue over weeks and months?

The answer developed by exercise physiologist Dr. Eric Banister in the 1970s and 1980s is a model built on three metrics: ATL, CTL, and TSB.

ATL — Acute Training Load

ATL represents your recent training load, typically calculated as a 7-day exponentially weighted average of Training Stress Scores (TSS). It reflects how fatigued you currently are. High ATL means you've been training hard recently and are carrying fatigue.

CTL — Chronic Training Load

CTL represents your longer-term training load — a 42-day exponentially weighted average. It approximates your fitness level. Higher CTL means you've built a stronger aerobic base. CTL increases slowly through consistent training and decreases during periods of reduced training or illness.

TSB — Training Stress Balance

TSB is simply CTL minus ATL. It's the balance between your fitness and your fatigue. When TSB is significantly negative (say, below -20), you're fatigued relative to your fitness and are at higher injury risk. When TSB is near zero or positive, you're fresh and ready to perform.

The art of marathon training — and the science that AI coaching applies — is managing these three numbers across a 16–20 week training block.

Periodisation: the structure of a marathon build

Periodisation is the organised structuring of training into phases, each with different objectives. Traditional marathon plans use linear periodisation: base building (weeks 1–6), build phase (weeks 7–12), peak (weeks 13–14), taper (weeks 15–16).

An AI marathon coach applies adaptive periodisation — the same structure, but with the boundaries and intensity targets adjusted based on your actual training data rather than fixed week numbers.

If your CTL is building faster than expected (you're adapting well), the system moves you into the build phase earlier. If your CTL is lagging (you've missed sessions, had illness), it extends the base phase rather than pushing you into intensity work your aerobic system isn't ready for.

Pace zones: the 80/20 principle

Elite marathoners spend roughly 80% of their training at easy effort (Zone 1–2) and 20% at moderate-to-hard effort (Zone 3–5). This ratio has been observed across multiple endurance sports and is the basis of polarised training theory, developed by sports scientist Dr. Stephen Seiler.

Most recreational runners get this backwards. They run most of their miles at a moderate "comfortably hard" pace that's too easy to provide strong adaptation stimulus but too hard to allow full recovery. The result: accumulated fatigue without corresponding fitness gains.

A good AI marathon coach monitors your pace zone distribution and nudges you toward the optimal balance. Easy runs should genuinely be easy. Hard sessions should be hard. The middle ground should be rare.

How injury risk detection works

The most powerful application of training load science is injury risk detection. Most overuse running injuries — stress fractures, IT band syndrome, plantar fasciitis — are preceded by identifiable patterns in training data:

• Sudden mileage increases: Jumping more than 10% in weekly volume week-over-week significantly increases injury risk
• High TSB deficit: Sustained periods of TSB below -30 indicate accumulated fatigue that the body can't absorb
• Elevated easy-run heart rate: If your HR on easy runs drifts up over consecutive weeks, it signals your body is under stress beyond what the pace data shows
• Insufficient recovery between hard sessions: Hard sessions less than 36 hours apart don't allow adequate muscle repair

An AI marathon coach monitoring these signals can flag risk 1–3 weeks before you'd typically feel symptoms. That's the difference between adjusting your plan and spending 6 weeks in a boot.

How session targets are set

Every session in an AI-generated training plan has purpose-driven targets derived from your current fitness data:

• Easy run pace: Set from your estimated aerobic threshold — the pace you can sustain while holding a conversation without effort
• Tempo pace: Your lactate threshold pace, estimated from recent race performances or time trials
• Interval targets: VO2max pace, calibrated to your current fitness level
• Long run pace: Typically 45–90 seconds per mile slower than your goal marathon pace

These targets update as your fitness changes. If your 5K time improves by 30 seconds mid-training block, your tempo and interval targets should update accordingly. A static plan can't do this. An AI marathon coach does it automatically.

The taper: science and psychology

The taper — the 2–3 week reduction in training volume before your race — is one of the most misunderstood phases of marathon preparation. Runners often feel awful during taper (the "taper madness" phenomenon) and respond by running too much, arriving at the start line fatigued rather than fresh.

Good AI coaching manages the taper using TSB targets: the goal is to raise your TSB to approximately +10 to +20 by race morning. Too fresh (+30 or above) and you've lost fitness. Too fatigued (below 0) and you've undertaken the race at a disadvantage.

This mathematical approach to taper management removes the guesswork that trips up so many self-coached runners in the final weeks.

The honest limitations

AI marathon training plans are built on data. The more data you provide — run history, Strava integration, heart rate, subjective feel ratings — the more accurate the model becomes. A new runner with no history gets a competent generic plan. A runner with 6 months of Strava data gets a plan that's genuinely calibrated to them.

What AI can't yet do is observe your running form, notice that you're limping slightly on your left side, or pick up on the qualitative signals that an experienced human eye can read. For most recreational runners, this limitation is acceptable. For runners with recurring biomechanical issues or complex injury histories, occasional in-person coaching remains valuable alongside AI-assisted training.

The science is real. The adaptation is genuine. Understanding how it works helps you trust the plan on the hard days — and know when to override it.