‘The algorithm and the acre’: Why the future of potato breeding will still be decided in farmers’ fields

A new global potato-breeding initiative promises faster climate resilience—but the real test will still happen in farmers’ fields

By Lukie Pieterse, Editor/Publisher of Potato News Today

On 18 August 2026, the International Potato Center announced a partnership that intends to use artificial intelligence to help breeders identify stronger potato varieties sooner. The promise is considerable. So is the need for a clear-eyed understanding of what the technology can—and cannot—do.

The new initiative is called GAIN-RT, short for Genetic Acceleration through Artificial Intelligence for Root and Tuber Crops. It brings together the International Potato Center (CIP), the James Hutton Institute in Scotland, the Kenya Agricultural and Livestock Research Organization (KALRO) and Egerton University. It was commissioned by the UK-CGIAR Centre, whose secretariat is CABI, and is funded by UK International Development.

At first glance, the announcement contains all the language that can make a practical potato person cautious: artificial intelligence, advanced models, genomic datasets and predictive breeding. Yet beneath the fashionable terminology is a serious proposition.

Potato breeders generate more information than any individual can weigh consistently—genetic markers, disease scores, weather records, soil data, canopy images, yield results, tuber-quality measurements and observations from different locations and seasons. A well-designed model may help them find useful patterns earlier and discard weak material before more years and money are invested in it.

That is worth exploring. But it would be a mistake to imagine a computer designing a perfect potato and sending it straight to a grower. Breeding remains biological, local and stubbornly dependent on time. The algorithm may help decide where to look. The acre still decides whether the prediction survives contact with reality.

A project with a clear starting point

GAIN-RT will initially focus on potato, with sweetpotato and other crops expected to follow. According to the CIP launch announcement, the partners intend to combine genomic, environmental and field data to predict which breeding material is most likely to perform under climate stress, carry more durable disease resistance and justify further testing.

CIP brings its breeding programmes and links with national systems. The James Hutton Institute contributes expertise in artificial intelligence, genomics and plant science. KALRO and Egerton provide regional knowledge, field validation and routes towards adoption in Africa.

CIP Director General Simon Heck put the urgency plainly: “Climate change and pests do not wait. They evolve rapidly. Artificial intelligence gives us the tools to catch up.”

That is the central claim behind the initiative. If prediction improves, breeders should be able to make better selections earlier, test more promising material across more environments and direct scarce field capacity towards the crosses most likely to succeed.

The launch statement does not specify the project’s budget, duration, numerical targets for shortening breeding cycles or the rules that will govern shared data. Those details will matter. For now, GAIN-RT should be seen as a significant research commitment, not as proof that the difficult part has already been solved.

What ‘AI breeding’ actually means

In this setting, artificial intelligence is not a chatbot writing variety descriptions. It is a family of statistical and computational methods used to estimate how untested or partly tested plant material may perform.

Models can be trained on known relationships between DNA markers, observed traits and environmental conditions. They can also use images collected by drones or ground-based cameras to describe crop development and differences among trial sites.

The attraction is straightforward. Conventional potato breeding produces thousands of seedlings and clones, while only a very small proportion can be carried through repeated field trials, quality evaluation, disease screening and seed multiplication.

Every poor selection that can be removed with reasonable confidence at an earlier stage creates space for a better candidate. Just as importantly, predictive tools can help breeders choose crosses that combine useful traits rather than relying solely on what performed well in the last familiar environment.

The word predictive is important. A model does not discover truth; it estimates probability from the information it has been given.

If its training data come mainly from irrigated, high-input sites, it may be less reliable in rainfed smallholder fields. If local varieties, women farmers’ preferences or informal market requirements are poorly represented, the model will not magically correct the omission. It will learn the limits and biases of the dataset with impressive mathematical confidence.

Potato remains a difficult customer

Potato is not the easiest crop on which to demonstrate rapid predictive progress. Most commercial cultivars are tetraploid and highly heterozygous. Many are related through repeated use of successful parents, while clonal propagation preserves both desirable combinations and accumulated disease risk.

Yield and quality are also shaped heavily by the interaction between genetics and environment.

The traits that matter are rarely independent. A clone may resist late blight but mature too late. It may tolerate drought yet fail a processor’s dry-matter or fry-colour specification. It may yield impressively but bruise, sprout early in storage, produce undesirable tuber shape or be rejected because seed multiplication is too slow.

Breeders are not choosing a single winning number; they are managing a crowded set of compromises.

A 2024 study of genomic prediction in tetraploid potato illustrates the unevenness. Researchers evaluated 762 offspring derived from 18 elite cultivars. Reported cross-validated prediction correlations reached 0.75 for dry-matter content, but only 0.28 for yield.

That does not make genomic prediction a failure. It shows why broad claims about “accelerating breeding” must be translated into trait-by-trait evidence.

Promising evidence—and an honest warning

There are good reasons for qualified optimism.

In July 2026, researchers Alexandre Hild Aono and Aakash Chawade reported a phenomics-assisted sparse-testing approach for potato breeding. Their work used image-derived information to characterise trial environments and improve predictions across three Swedish environments.

In several scenarios, the image-based environmental model performed as well as, or better than, models relying on genomic information alone. The strongest results were obtained for tuber yield.

Sparse testing allows a breeding programme to evaluate more genotypes without planting every candidate at every location. The missing observations are estimated from relationships among genotypes, environments and measured traits.

If prediction is dependable, the same budget can expose more material to more relevant conditions—an especially useful prospect for public breeding programmes with limited land, staff and genotyping capacity.

The authors were careful about the limits. They called for stricter validation and testing in independent, more strongly contrasting environments.

That caution should be welcomed, not buried. Climate resilience is precisely the area in which a model trained on yesterday’s conditions may face tomorrow’s surprise.

Climate resilience cannot be reduced to one score

CIP reports that potato occupied approximately 17.1 million hectares and produced 390.4 million tonnes globally in 2024. Behind those large figures are sharply different production systems: mechanised irrigated farms, highland smallholdings, seed plots, processing contracts, fresh markets and household production.

The “best” climate-resilient potato will not be the same everywhere.

I have spent enough years around potato growers and breeders to know that a new variety is never merely a collection of superior genes. It must fit a season, a market, a kitchen or factory, a storage system, a seed supply and a farmer’s tolerance for risk.

Heat, drought, late blight and bacterial wilt matter, but so do taste, texture, dormancy, tuber appearance, cooking behaviour, processing recovery and the price at which clean seed can be obtained.

Artificial intelligence can help compare these variables, but people must decide which ones deserve priority. That means target product profiles must be built with growers, seed producers, traders, processors and consumers, not written only by data scientists or donors.

Otherwise, a project can optimise traits that are scientifically measurable while missing the reasons farmers actually adopt—or refuse—a variety.

Africa must be a partner, not a testing ground

The inclusion of KALRO and Egerton University is therefore more than a matter of geography.

Local scientists and field teams understand disease patterns, production constraints, seed channels, market preferences and the practical meaning of risk in ways that cannot be imported with software. Their role should extend from supplying trial data to shaping the questions, owning analytical capacity and helping determine how benefits are shared.

GAIN-RT also plans to train scientists, with particular attention to women and early-career researchers. This may prove to be one of its most durable contributions.

A breeding platform that depends indefinitely on outside analysts or inaccessible computing infrastructure is not genuine capacity development. Local teams need the skills and authority to challenge a model, retrain it and decide when its recommendation does not fit what they see in the field.

Data governance deserves equal attention. Farmers and national programmes should know what information is collected, who can use it, where it is stored and whether resulting tools or varieties will remain affordable.

The launch announcement does not yet answer those questions. They should be addressed while the partnerships are being built, rather than after valuable datasets have changed hands.

A released variety is not yet an impact

The day before GAIN-RT was announced, CIP highlighted recent variety releases that show what the existing breeding network is already delivering.

Ethiopia released eight potato varieties in 2026, including material aimed at drought-prone areas, acidic soils, irrigation and processing markets. Under favourable conditions, Beri and Midekisa were reported to yield as much as 53 and 50 tonnes per hectare respectively.

Malawi released four varieties combining yield potential, late-blight tolerance, tuber size and processing quality.

Mozambique released five varieties after evaluation in several agro-ecological zones with the participation of more than 500 farmers. Reported yields ranged from 22 to 42 tonnes per hectare, with attention given to pest and disease resistance, including bacterial wilt. CIP’s account of the releases provides an encouraging picture of what coordinated breeding programmes can deliver.

These are meaningful achievements. They also reveal the next bottleneck.

A variety can be officially released and remain largely absent from farmers’ fields. Early-generation seed must be multiplied, disease health protected, certification made workable, commercial seed growers engaged and demand created.

Traders and processors must recognise the product. Farmers must be able to obtain seed at the right time and at a price that does not transfer all the risk to them.

No prediction model performs those tasks. If GAIN-RT accelerates selection but the seed system does not accelerate with it, genetic progress will wait in a catalogue.

What the industry should watch

The most useful measure of GAIN-RT will not be the number of algorithms developed or genomic records processed.

It will be whether candidate varieties reach independent multi-location testing sooner; whether prediction holds under unfamiliar weather and disease pressure; whether breeders retain enough genetic diversity rather than repeatedly selecting the safest familiar material; and whether farmers receive seed of varieties they helped define.

Transparency will help. The partners should publish performance benchmarks, including where models fail. They should explain how much time and field cost is actually saved for different traits, how predictions transfer between regions, and how farmer preferences are incorporated.

Negative results are especially valuable in a field now crowded with optimistic claims about artificial intelligence.

A 2026 review of genomics-assisted breeding in root and tuber crops reached a similarly practical conclusion: computational advances must be accompanied by trained multidisciplinary teams, accessible data infrastructure, multi-environment trial networks, participatory breeding and workable delivery systems.

The announcement gives reason for interest, but not for surrendering judgement. That is not cynicism. It is the kind of discipline that protects good science from becoming a fashionable demonstration project.

The field remains the judge

GAIN-RT arrives at a moment when potato breeding genuinely needs more speed.

Weather patterns are shifting, pests and pathogens continue to evolve, and growers are being asked to produce reliable crops with less water, fewer crop-protection options and tighter margins. Better prediction could help public and private breeders use their limited years, plots and people more intelligently.

But the most important word in predictive breeding is not artificial. It is breeding.

Progress still depends on good parents, clean measurements, representative trials, skilled technicians, honest failure, patient seed multiplication and farmers willing to test something unfamiliar. Technology strengthens that chain only when it respects every link.

A computer may help tell a breeder where to look. It cannot decide what a good potato means in a particular place, or carry the consequences when a supposedly resilient variety disappoints a family, a seed grower or a processor.

The acre keeps that authority.

If GAIN-RT remembers this, the project could become an important step towards faster and more relevant potato improvement. If it forgets, it may produce impressive predictions while the crop—and the people who depend on it—move on without them.