A year ago, GPT-3 was mostly something people saw in impressive demos on Twitter. Today it sits behind hundreds of applications through OpenAI's API. In June, GitHub launched a technical preview of Copilot, which suggests whole functions inside your code editor using OpenAI Codex, a descendant of GPT-3 trained on public code. In May, Google showed LaMDA, a conversational model, at its developer conference. And last month, more than a hundred researchers at Stanford published a 200-page report arguing that these systems deserve a name of their own: foundation models.

At the same time, the most-discussed AI ethics paper of the year asked a blunt question: can language models be too big? The controversy around that paper, and the departures from Google that accompanied it, turned a technical debate into a debate about power: who builds these systems, who studies their risks, and who decides when they are ready.

This post summarises the main ethical concerns about large language models as they stand in 2021, the regulatory response taking shape in Europe, and what teams building with these models can do today.

1. A year of scale

Date Event
June 2020 OpenAI announces GPT-3, a 175-billion-parameter language model, and opens a limited API
December 2020 Timnit Gebru, co-lead of Google's Ethical AI team, leaves Google after a dispute over a paper on large language models. She says she was fired; Google says it accepted her resignation
January 2021 Researchers from Stanford and McMaster University publish a study showing persistent anti-Muslim bias in GPT-3
February 2021 Google fires Margaret Mitchell, the other co-lead of the team. Google says she violated its code of conduct and security policies
March 2021 The paper, "On the Dangers of Stochastic Parrots", is presented at the ACM FAccT conference
April 2021 The European Commission proposes the Artificial Intelligence Act
May 2021 Google presents LaMDA at Google I/O
June 2021 GitHub announces Copilot, powered by OpenAI Codex
August 2021 Stanford's Center for Research on Foundation Models publishes "On the Opportunities and Risks of Foundation Models"

What links these events is scale. The same property that makes these models so capable, training on enormous amounts of text from the internet, is the source of most of the ethical concerns.

2. What the "Stochastic Parrots" paper actually argues

The paper by Emily M. Bender, Timnit Gebru, Angelina McMillan-Major and Margaret Mitchell (listed under a pseudonym) is often summarised as "AI is biased". Its argument is more specific. It names four groups of risks:

Risk The argument
Environmental and financial cost Training ever-larger models uses large amounts of energy and money. The costs fall on everyone, including communities affected by climate change, while the benefits go mostly to well-resourced organisations and languages
Unfathomable training data Web-scale datasets are too large to document or audit. They over-represent some voices (younger, wealthier, English-speaking, online) and encode stereotypes and abusive language. Filtering often removes the voices of marginalised groups along with the toxic content
Research direction Investment in ever-larger models draws attention away from approaches that try to build genuine understanding, and from smaller, better-documented models
Fluent text without meaning Language models produce text that sounds coherent but is not grounded in any intent or understanding. Humans naturally read meaning into fluent text, which makes the output persuasive even when it is wrong or harmful

The memorable phrase is in that last point. A language model, the authors write, is a system for stitching together sequences of linguistic forms it has observed in its training data, "without any reference to meaning: a stochastic parrot".

You do not have to agree with every claim in the paper to see why it matters. Its central recommendations are practical and hard to argue against: document your datasets, weigh costs before building bigger, involve affected communities, and evaluate risks before deployment, not after.

3. Bias is measurable, and it is not subtle

Arguments about bias in language models can sound abstract until you look at the data. In January, Abubakar Abid, Maheen Farooqi and James Zou tested GPT-3 with prompts such as "Two Muslims walked into a". In 66 percent of completions, the model produced text involving violence. The rate was far lower for other religious groups. When asked to complete analogies, GPT-3 mapped "Muslim" to "terrorist" in 23 percent of cases.

The researchers tried to reduce the effect by adding positive adjectives to the prompt. The best result brought violent completions down to around 20 percent, still higher than for other groups. The bias was not a quirk of one prompt. It was persistent.

This is the core problem with web-scale data. A model trained to predict what text is likely will reproduce what is common in its training data, including stereotypes. And because the output is fluent and confident, it can be harder to spot than an obviously broken system.

4. Who gets to ask the hard questions?

The controversy around the paper raised a second question, separate from the technical one: can ethics research be independent when it is done inside the companies building the technology?

The details of Gebru's and Mitchell's departures are disputed, and both sides have given their accounts in public. What is not disputed is the effect. Thousands of Google employees and outside academics signed letters of protest, and some researchers have since turned down Google funding. The episode made it clear that corporate AI ethics teams can face a conflict of interest when their findings threaten a core business.

That does not mean in-house ethics work is worthless. Companies have access to data, systems and decision makers that outside researchers do not. But it does suggest three things:

  • Independent research matters. Universities, non-profits and journalists need access to models and data to study them.
  • Process matters. How a company reviews and approves research that might be critical of its products says more about its commitment than its published principles.
  • Concentration matters. Only a handful of organisations can afford to train the largest models. That gives them a large influence over what gets built, what gets studied, and what counts as "responsible".

5. Foundation models: a new name for a new kind of risk

The Stanford report, led by Rishi Bommasani, Percy Liang and more than a hundred co-authors, introduces the term foundation model for models trained on broad data at scale that can be adapted to many downstream tasks. BERT, GPT-3 and CLIP are examples.

The authors chose the word to emphasise that these models are central but incomplete. Many applications are built on top of the same few models. That creates two linked properties:

Property Meaning Consequence
Emergence Capabilities arise from scale rather than being explicitly designed Builders often do not know in advance what a model can do, or how it will fail
Homogenisation Many systems are built on the same base model Improvements spread quickly, but so do flaws: any bias or weakness in the base model is inherited by everything built on it

The report has its critics, some of whom argue that the name gives these models more credit than they deserve. But the homogenisation point is important for anyone building products. If you build on a foundation model, you are inheriting decisions about data and training that you did not make and may not be able to inspect.

6. The regulatory response: Europe's proposed AI Act

In April, the European Commission published its proposal for an Artificial Intelligence Act, the first attempt at comprehensive AI regulation by a major jurisdiction. It takes a risk-based approach:

Risk level Examples in the proposal Treatment
Unacceptable Social scoring by public authorities, manipulative techniques that exploit vulnerabilities, most real-time remote biometric identification in public spaces for law enforcement Prohibited, with limited exceptions
High AI used in hiring, education, credit scoring, access to public services, law enforcement, critical infrastructure Allowed, with requirements on data quality, documentation, transparency, human oversight and accuracy
Limited Chatbots, deepfakes Transparency duties: people should know they are interacting with an AI or seeing synthetic content
Minimal Spam filters, video games No new obligations

Proposed fines for the most serious violations go up to 6 percent of global annual turnover or 30 million euros, whichever is higher.

The proposal is only the beginning. It now goes to the European Parliament and the Council, and it will likely take years to agree. One open question is how it will deal with general-purpose models like GPT-3, which are not built for a single use and so do not fit neatly into a list of high-risk applications. The risk often depends on what someone downstream does with the model.

7. What teams building with language models should do now

You do not have to wait for regulation to act responsibly. If you are building a product on top of a large language model, these practices are within reach today.

Before you build

  • Ask whether you need a large model at all. A smaller, well-understood model or a rules-based system may be safer and cheaper for narrow tasks.
  • Map who could be harmed. Think about users, people mentioned in outputs, and people affected by decisions the system supports.
  • Read the model card or documentation. If there is none, treat that as a risk.

While you build

  • Test for bias explicitly. Run prompts that vary names, religions, genders, nationalities and dialects, and compare the outputs. Do not assume the base model has been fixed.
  • Keep a human in the loop for consequential decisions. Language models should not decide alone on hiring, credit, health or legal matters.
  • Filter, but check what you filter. Toxicity filters can disproportionately block dialects and identity terms used by minority groups.
  • Design for wrong answers. Fluent text is not correct text. Show sources where possible, and make it easy for users to flag errors.

After you launch

  • Monitor real use. People will use the system in ways you did not plan for.
  • Document decisions. Record what model you use, how you tested it and what limits you set. If regulation arrives, you will need this anyway.
  • Create a path for concerns. Make it safe for employees and users to raise problems, and act on them.

Why this matters

  • The models are already in products. These are no longer research questions. Code assistants, writing tools and chatbots built on large language models are being used today.
  • Bias at scale is still bias. A model that reproduces stereotypes in a fraction of outputs will do so millions of times once deployed.
  • Power is concentrated. A few organisations train the models everyone else builds on. Their choices, and the independence of the people who study them, affect everyone.
  • Regulation is coming. The EU proposal shows the direction: risk-based rules, documentation, transparency and human oversight. Teams that start now will be ready.

Final Thoughts

The debate about large language models is often framed as optimists versus pessimists: people who see a path to remarkable new tools and people who see risks being ignored. That framing misses the point. Both can be true. These models are genuinely capable, and they genuinely carry risks that grow with their scale and reach.

The more useful question is not whether to build them, but how decisions about them are made: who is at the table, what evidence is required before deployment, and whether the people raising concerns are heard or pushed out. This year has shown that the answers are still being worked out, inside companies, in research communities and in parliaments. Anyone building with these models is now part of that conversation, whether they like it or not.

Sources