Foundation models can make it easier to build many applications from a shared, broadly trained model. That reuse can accelerate experimentation and support work in fields such as science, healthcare, law and education—but it can also carry shared weaknesses, privacy and security problems, and biased behavior into systems people rely on. Whether a particular model is useful or safe depends on the task, the people affected and the safeguards around its deployment; broad benchmark performance alone is not enough to establish reliability.
What is a foundation model?
Stanford’s Center for Research on Foundation Models (CRFM) describes a foundation model as one trained on broad data, generally through self-supervised learning at scale, and adaptable to a wide range of downstream tasks. Developers can adapt a shared model—for example, by fine-tuning it—instead of building and training a separate model from scratch for every application.
“Foundation model” is not another name for “generative model.” Some foundation models generate content, while others may classify or otherwise process inputs; not every generative or discriminative model meets the foundation-model definition. The important feature is broad pretraining followed by adaptation to varied tasks.
Reuse creates leverage: one base model can support many downstream systems. It also creates a route for shared defects to spread. An adapted model is not automatically suitable for its new purpose simply because the base model performed well elsewhere.
#1 Best Overall
What opportunities do foundation models create?
Lower barriers to experimentation
Using a pretrained model can spare a downstream developer from assembling a large training dataset and paying to train a foundation model. It may make it possible to explore an application that would otherwise be out of reach. The savings are not complete: integration, compute, data handling, evaluation and ongoing oversight still take resources, and developers may depend on a model provider or infrastructure they do not control.
Productivity and scientific work
The OECD identifies productivity gains and faster scientific progress as potential benefits of AI. A foundation model may help with particular research or work tasks, but those are possibilities, not guaranteed outcomes across occupations or fields. The useful question is whether it improves a defined task under realistic conditions, without shifting hidden costs or errors onto workers and users.
Applications across fields
- Healthcare: Models may support interfaces, biomedical research, or tasks involving text, images and molecules. Biased datasets and inadequate trials can undermine usefulness or produce unequal outcomes, so domain-specific validation matters.
- Law: Generative tools may assist with drafting. That does not establish dependable legal reasoning, accurate statements of fact or sound use of sources; factuality and provenance remain important concerns.
- Education: Interactive feedback and personalization are possible uses. Their value depends on the model’s capabilities in the subject and on whether it has been adapted and evaluated responsibly for learners.
These examples illustrate potential uses, not proof that a foundation model is appropriate for every high-stakes deployment in those fields.
Why does reuse also create risks?
Risks can arise in the model itself and in the way a downstream application is designed or used. Stanford CRFM distinguishes intrinsic bias in a model from extrinsic harm in a particular application. A biased training dataset or design choice may shape a model’s behavior; deployment decisions determine who encounters that behavior and what consequences follow. Identifying the source of a harm is important for deciding who can prevent or correct it.
Inherited bias and inequity
When many applications rely on the same base model, its limitations can travel with them. Adaptation may change behavior, but it does not by itself establish that bias has been removed. Developers need to evaluate relevant populations and use cases rather than assume that results for one group or task apply to another.
Unreliable answers and evaluation gaps
Strong results on broad benchmarks do not establish truthful, robust performance in a consequential setting. Models can fail in ways that are difficult to predict, especially when real inputs differ from evaluation data. In applications such as legal work, the ability to check factuality and trace claims to sources is a particular concern. Representative testing and clear review procedures are more informative than a general capability claim.
Rank #3
Privacy and security exposure
Foundation models may memorize parts of their training data, and they can be vulnerable to adversarial attacks. General-purpose capabilities may also be used in unintended ways. Organizations should assess what sensitive information users may submit, who can access the model and its outputs, how data is retained, and what controls are in place to detect and respond to abuse.
Misuse and information harms
Models can lower the effort involved in producing targeted disinformation or deepfakes used for harassment. The OECD also identifies manipulation, disinformation, fraud and cyberattacks as prospective AI risks. The scale and form of harm depend on the model’s capabilities, how accessible it is and how it is used; these possibilities should not be treated as inevitable outcomes of every deployment.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Environmental costs
Training foundation models can require substantial computation and energy. A meaningful environmental assessment also considers inference use, energy sources, hardware and what the model is being compared with. Without those details, a training figure alone cannot establish the full footprint. Stanford CRFM calls for better documentation and measurement of these costs.
Concentration and dependence
The expense and complexity of developing foundation models can favor well-capitalized companies and governments, concentrating ownership and power. Pretrained models can lower some barriers for downstream developers, but that does not remove reliance on providers, infrastructure or deployment expertise. OECD-reported investment figures illustrate the growth of financing in a specific period: global venture-capital investment in AI startups rose from USD 31 billion in 2015 to USD 98 billion in 2023. Generative AI’s share of total AI venture-capital investment rose from 1% (USD 1.3 billion) in 2022 to 18.2% (USD 17.8 billion) in 2023. These figures describe investment over those periods; they are not present-day market totals or evidence that the technology has delivered productivity gains or that benefits outweigh risks.
Legal and governance uncertainty
Questions about liability, data rights, transparency and model release remain active policy issues. The OECD identifies clearer liability rules and risk management as policy priorities. An organization adopting a model still needs to determine which legal and compliance obligations apply to its own use.
What does “open weights” mean?
In its 2025 primer, the OECD defines open-weight models as foundation models whose trained weights are publicly available for download for local deployment. This describes access to weights; it does not, by itself, establish that training data is transparent, the license permits a particular use, the model is safe, or that users can practically modify and deploy it. The OECD primer does not cover licensing, while noting that licensing remains a critical deployment factor.
Best Value
Choosing between a hosted service and downloadable weights is therefore a question of control and responsibility, not a simple open-versus-closed ranking. A hosted service may leave important data, update and access decisions with a provider; local deployment can give an organization different forms of control, but also makes its infrastructure and operational responsibilities central. Evaluate the actual service terms or model license rather than infer them from the access label.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How should an organization evaluate a foundation model?
NIST’s Generative AI Profile is a voluntary, cross-sector companion to AI RMF 1.0. It is intended to help organizations incorporate trustworthiness considerations into the design, development, use and evaluation of AI products, services and systems. It is a risk-management aid, not a certification or guarantee of safe outcomes. The following questions turn recurring technical, social and governance concerns into a practical comparison:
- Access and control: Is the option a hosted API or downloadable weights? Check data residency, update control and dependence on a provider.
- Task evidence: Test representative tasks and populations. Consider distribution shifts, the severity of possible errors and whether evaluation is independent.
- Data and rights: Examine the provenance and suitability of training and prompt data, privacy exposure, retention practices and applicable licensing terms.
- Security and misuse: Assess access controls, monitoring, adversarial testing and the process for responding to abuse.
- Deployment context: Identify who may be affected and how serious errors could be. Decide where human review is needed, whether people can appeal, and how errors can be corrected.
- Cost and footprint: Account for total training or inference costs and energy use, and compare the model with smaller models or non-model alternatives.
- Governance: Assign accountability, document risk reviews and compliance obligations, and establish a process to monitor changes over time.
For consequential uses, the decision should turn on evidence from the intended task and safeguards that match the possible harm—not on general claims about a model’s breadth. If the organization cannot evaluate or correct likely failures, a narrower tool or non-model approach may be the more responsible choice.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
