“Is artificial intelligence the new doping?” asked former Movistar director of performance Mikel Zabala at this year’s Science & Cycling conference, held each year the week prior to the Tour de France Grand Départ. It’s where cycling’s most progressive academics, sport scientists and WorldTour staff present their work, featuring the latest research on training, nutrition and equipment.

But Zabala’s presentation stood out as the experienced Spaniard warns of parallels between the Armstrong era and the artificial era. Why? We’ll come back to Zabala’s psychosocial reasoning later. But Zabala did shine a spotlight on AI and why machine learning could be an anti-doping scientist’s worst nightmare…

Creating new drugs

March 2025 could prove a pivotal date in the world of anti-doping, as the United States Adopted Names (USAN) Council granted biotech company Insilico Medicine’s AI-designed drug to treat idiopathic pulmonary fibrosis (IPF) an official name: Rentosertib, formerly known as ISM001-055.

Rentosertib marked a significant milestone in drug development: its target and molecule were identified and designed using generative AI, taking an AI-created drug into human trials. For the cycling chemists amongst you, Insilico used its Pharma.AI platform to accelerate the discovery of Rentosertib. That followed PandaOmics, an AI-powered biology engine, identifying the enzyme TNIK as a promising target for idiopathic pulmonary fibrosis. Finally, the platform’s generative chemistry engine, Chemistry42, designed and optimised the compounds against it, ultimately leading to Rentosertib’s selection as a pre-clinical candidate.

It’s impressive stuff and, for cycling, the implications are stark. If AI can rapidly design novel medicines, could it eventually create performance-enhancing substances that anti-doping laboratories have not yet detected?

“We haven’t seen that yet,” an anti-doping source told us, who wished to remain anonymous, “but we’ve certainly seen existing drugs, like steroids, altered by AI for more specific performance benefits. It is a concern. And potentially dangerous.”

Cycling has already seen athletes caught using experimental drugs such as GW1516. GW1516 is known by various names, including cardarine and endurobol, and was designed as a dietary supplement. However, it was pulled at the experimental phase because it was found to cause cancer. Despite that, it made its way to the black market via underground labs, which is why WADA banned it.

Those toxicity concerns didn’t deter several cyclists from experimenting with the drug, including Russian rider Valery Kaykov, who was part of the nation’s successful team pursuit at the 2012 European Track Championships, the same year he joined Pro Continental team RusVelo. A year later, he became the first athlete to officially test positive for GW1516 and was subsequently sacked by his team.

Ostarine is another drug that was never approved as a medicine but found its way into the sporting arena, with athletes tempted by its potential to support muscle mass. Italian rider Matteo Spreafico tested positive for enobosarm (ostarine) at the 2020 Giro d’Italia and received a three-year ban. His compatriot Linda Laporta returned an adverse analytical finding for the same substance at this year’s UAE Tour Women and was provisionally suspended pending the outcome of her case, which is ongoing, and has not resulted in a confirmed anti-doping rule violation.

With AI now capable of designing novel drug candidates, the next generation of performance-enhancing substances could potentially emerge faster than anti-doping science can identify them. AI has the capacity to take ‘unknown science’ and make drug development much more efficient and, seen through an anti-doping lens, arguably harder to detect.

AI could also enable more personalised enhancement by analysing physiological and genetic data to identify interventions tailored to an individual athlete. This could shift doping from simply finding effective drugs to identifying the most effective intervention for each athlete, creating an increasingly sophisticated challenge for anti-doping authorities.

Mimicking for good and bad

“This extends to the idea of a digital twin,” says Professor Yannis Pitsiladis, director of the Centre for Exercise and Medicine at Hong Kong Baptist University. He’s also on WADA’s health medical research committee.

“A digital twin is gaining traction in the medical world,” he adds. “In essence, it’s a virtual model of an individual that uses real-world data to simulate health, disease and treatment responses. It might combine genetic information, blood tests, medical history, imaging, medications, activity, heart rate, sleep, nutrition and other wearable data. Doctors could use it to predict outcomes and personalise treatment.”

In cycling, a digital twin would be fed similar data, so power, heart rate, cadence, speed, sleep, training load, nutrition, environmental conditions and potentially physiological/genetic information to create a model that behaves like the individual rider. The key is that it’s continually updated with real-world data from the rider.

In cycling, a sophisticated digital twin could potentially answer questions such as: ‘What happens if I increase training volume by 10%?’ ‘How will riding at 2,000 metres affect my performance?’ and ‘What nutrition strategy is likely to work best on a hilly parcours at 10°C?’

Rather than simply reporting what happened, the digital twin attempts to predict what will happen next.

The idea of digital twins has already hit the WorldTour. At the start of 2026, Ineos Grenadiers (now Netcompany Ineos) and Swansea University revealed they were developing digital twins to analyse riders’ performance and development, aiming to identify overlooked talent and predict which young cyclists could become future professionals. The legal benefits of a digital twin are clear. But the same technology could theoretically personalise doping – identifying substances or strategies with maximal benefit and minimal detectable impact.

“Of course, if you have enough data, you might be able to ask what the best combination of testosterone, erythropoietin (EPO) and cortisone might be, and at what dose, for what specific outcome,” says Pitsiladis. “But you’d need good data.”

Of which cycling has. In fact, cycling’s so obsessed with data that even metabolomics is being used to study elite cyclists, including at UAE Team Emirates-XRG when Professor Iñigo San Millán was heavily involved with the team. Metabolomics involves measuring hundreds or thousands of small molecules (metabolites) in blood or other biological samples to better understand a rider’s metabolism for more accurate training prescription. It’s a level or thousand up from measuring a rider’s blood sample for iron before they train at altitude.

“With the amount of data cycling has, of course it could happen,” adds Pitsiladis, “but it could also benefit anti-doping by giving the agencies a clearer picture of the athlete.”

Strengthening anti-doping

As could the work of Dr Michael Skinnider, assistant professor at the Lewis-Sigler Institute for Integrative Genomics at Princeton University, who’s working on a chemical AI platform that focuses on discovering new doping agents.

“We’re developing an AI model that can predict the chemical structures of future performance-enhancing drugs,” he says. “The second part of our work is working with the mass spectrometry data [mass spectrometry machines analyse an athlete’s sample] to more accurately detect anomalies.

Typically, there are thousands – sometimes tens of thousands – of signals that might be associated with one small molecule. If there is suspicion that a novel drug is in this sample, it’s currently like finding a needle in a haystack. Our research aims to refine this work.”

Skinnider says the work is ongoing, but the signs are positive. “The first stage of the project involved curating a large database of 1,600 drugs that were either on the prohibited list or that we felt might have potential to enhance performance. It predicted two substances that were later added to the WADA banned list, plus it flagged up a number of novel substances of interest.”

Skinnider sees his work as a decision-support tool and, if it ever fulfils its potential and is adopted by WADA, it’d complement the biological passport. “If you saw some shift in a rider’s blood values and there’s talk that it might be down to some kind of novel performance-enhancing drug or substance, that’s where our work would come in.”

Avoid a return to the past

That’s the scientific AI battle. But what about the psychosocial? Let’s return to the start and Zabala’s concerns that AI will send cycling back to the dark ages, namely Armstrong, Puerto…

“We must never forget the past,” he says. “From Armstrong to Operation Puerto, cycling in the 1990s and 2000s was horrendous. It was a time when doctors ruled.”

Zabala calls this period and the time before, ‘Cycling 1.0’. He says that cycling followed a hierarchy and training relied on random coaching wisdom rather than empirical evidence. The absence of structured scientific collaboration blurred medical, ethical and professional boundaries.

“Then ‘Cycling 2.0’ took over in around 2012,” he adds. “This involved multidisciplinary co-operation between coaches, scientists, physicians and nutritionists. Now, experts were mainly in charge, not backroom staff who were dopers of the past. Importantly, the rider went from a passive subject – the knowledge is held by the doping doctor – to an informed collaborator. This enhanced compliance, motivation and accountability.”

In short, Cycling 2.0 – deemed cleaner than ever – passed ownership to the rider. We’re now onto ‘Cycling 3.0’ where AI can magnify potential through predictive modelling but, Zabala says, it must not replace human judgment. Over-reliance on AI without critical thinking can lead to poor decision-making and a loss of ownership.

“A rider must maintain their autonomy,” he says. “In the doping days, coercion and fear of exclusion was very real. With unregulated AI, we have similar fears. Biometric over-monitoring and opaque decisions could leave a rider vulnerable. Without ethical governance, AI could reverse 2.0’s gains.”

To that end, Zabala proposes that riders and coaches must become AI literate. “That means they capitalise on its benefits and not be captured by it. As we enter this third evolutionary stage, cycling stands at an unprecedented crossroads between human wisdom and machine intelligence. Artificial intelligence will not replace the coach, the scientist or the athlete, but it will augment them.

“The success of Cycling 3.0 will depend not on how quickly we adopt new algorithms, but on how wisely we integrate them into a culture of shared knowledge, respect, transparency and responsibility,” he says. “If the spirit of Cycling 2.0 remains our compass, then AI can become the next ally in a journey that should continue to be profoundly human. The challenge is not merely to predict performance but to protect meaning.”

Ultimately, like every sector of life, AI has the potential to transform cycling for better or worse. It may uncover talent, personalise training and strengthen anti-doping, yet potentially accelerate novel drugs and sophisticated enhancement. The lesson from Cycling 2.0 is clear: technology works best when athletes remain informed participants. Cycling 3.0 must embrace AI without surrendering human judgement, transparency or the hard-won lessons of its past.