Our approach

Why Bells Up AI Chose to Offer Models Some Lawyers Will Refuse to Use

Some lawyers will see a model in the Bells Up catalog from a company they refuse to support. Their objection may concern the company’s leadership or politics, its privacy practices or environmental impact, its business conduct, or the way its models were trained and the biases they may reproduce.

A provider is paid only when its model is used

The question that follows is fair: Why would Bells Up offer that model, and does subscribing to Bells Up put money in that provider’s pocket?

Here is the direct answer on money. Your subscription fee is paid to Bells Up and is not allocated among model providers. Bells Up has no fixed fee, minimum commitment, licensing fee, or other payment obligation arising from the availability of a model family. We incur a provider charge only when someone sends a request to that family. If your firm disables it, your firm’s subscription does not cause Bells Up to pay the service delivering that model.

Your firm controls the model list

An organization owner or administrator can disable an entire model family for every user in the firm. The family then disappears from the model list, preventing anyone at the firm from selecting it. An administrator can set these restrictions during signup, before any user accounts are created.

Individual users may hide additional model families from their own accounts. They cannot restore a family disabled by their organization.

In specific cases, Bells Up displays a notice when a model family is selected, summarizing particular concerns about that model.

There are valid reasons to use these models

We offer access because these models are already out in the market. Your client may consult one before contacting you. Opposing counsel may use one to develop an argument or draft a letter. A witness, expert, or vendor may rely on one. A legal or general-purpose AI product may incorporate one behind the scenes. You will encounter these models’ output even when you would never choose the models for your own work.

We view controlled access as a form of harm reduction. An investigative or comparative request incurs a provider charge just like any other request. The benefit is that you can examine how the model handles a question, compare its answers with those of other models, identify errors or bias, and prepare to challenge AI-assisted work from another source without relying on the model’s output in client work. If you or your firm objects to any use of the model or any resulting payment, you can disable the family entirely.

Different objections require different answers

Some objections concern the model itself. A lawyer may question its accuracy, bias, training, or handling of confidential information. Comparison, contractual protections, and data-path disclosures help a lawyer evaluate those risks.

Other objections concern the company behind the model. A lawyer may oppose its leadership, politics, environmental record, or business conduct. Examining the model will not resolve those concerns. Bells Up addresses them by giving firms the power to exclude the model family and by incurring no provider cost when the model is not used.

All models have an environmental impact; here’s how to reduce your footprint

Transparency is a core value for us, and we are proud of our extensive privacy disclosures. We are unable to achieve similar transparency regarding environmental claims because reliable, model-specific information about the environmental impact of the models we use is limited. Accordingly, Bells Up presents the available evidence and identifies material gaps rather than claiming more certainty than the evidence supports.

What you can do is straightforward even where the measurements are not. Use the smallest, least reasoning-intensive model that can complete the task reliably. If you are using Anthropic Claude, choose Claude Haiku when it can handle the task, and move to Sonnet or Opus when the task requires more capability; if you are using OpenAI GPT, choose Luna when it can handle the task, and move to Terra, where offered, or Sol when more capability is warranted. Reserve high-reasoning or always-reasoning models for work that benefits from their additional reasoning, because high-reasoning models can use substantially more energy per query. And remember that a model that produces an unusable answer may require additional queries, so a lower per-query estimate does not necessarily produce a lower footprint for the work as a whole.

Our companion paper, AI and Environmental Impact: What Bells Up Can—and Cannot—Measure, sets out the evidence and its limits in detail.

Comparison and verification

Comparing responses from different models can expose inconsistencies and identify issues that deserve closer review. A disagreement between models may reveal an area that requires greater scrutiny or further inquiry.

Every current-generation AI model can produce inaccurate or fabricated information. Lawyers remain responsible for checking any automated output they incorporate into their work.

The model company and the infrastructure company

The company that develops a model and the company that delivers it may be different. Some models are accessed directly from the developer. Others run through infrastructure operated by another company.

For directly accessed models, Bells Up incurs a charge from the model provider when a request is sent. For models delivered through Amazon Bedrock, Bells Up incurs a charge from Amazon when a request is sent. This distinction matters because a lawyer may object to the model developer, the infrastructure provider, or both.

Bells Up discloses the data path and identifies the parties that may handle customer data for each model. A firm that objects to an infrastructure provider should review the data path for every model it permits, since the same infrastructure provider may support several model families.

My responsibility

I am responsible for deciding which models Bells Up will offer and under what conditions. Bells Up evaluates whether each model can be made available with appropriate privacy and confidentiality protections. In some cases, we limit a model family’s functionality because a broader implementation would create an unacceptable privacy risk.

Providers, contractual terms, infrastructure, and available evidence will change. We regularly review the models we offer and update our disclosures. We will remove a model that no longer meets our requirements.

I built Bells Up to give solo and small-firm lawyers access to the same range of capable tools available to larger firms. That includes the ability to examine tools they distrust and exclude companies they refuse to support. The decision belongs to the lawyer and the firm.