Lawyers won't trust any old AI

Despite their tendency for risk-aversion, lawyers are eager to embrace AI tools. The core concern comes down to trust.

Bethell at Travers Smith says trustworthy AI has two dimensions:

“Trusting that the data is secure, and trusting that the model will produce reliable answers.”

For Bethell, this means using enterprise-grade platforms that process data in line with the firm’s privacy policies, with the right information security and data retention controls in place. It also means recognising that model reliability is still evolving as AI systems improve.

According to our survey, 85% of legal professionals at large law firms are concerned about inaccurate or fabricated AI outputs, alongside fears around confidentiality and over-reliance.

Hallucinations are now materially more manageable than they were two years ago, says Nick Capell, IT Director for Gateley.

"This is due to improvements in foundation models, better techniques for grounding AI outputs in authoritative source material, and increased user education and clearer internal policies."

Bethell has seen a similar shift as newer models have improved the reliability of outputs.

“We can expect to keep seeing these jumps in capability and associated trust in the tools shift.”

But, for firms, improving model quality does not remove the need for professional review. Bethell says the focus remains on “instilling the right behaviours around checking accuracy”.

The importance of getting these answers right is front-of-mind for decision-makers.

Trustworthy AI is less about eliminating risk entirely and more about making risk visible, controllable, and professionally accountable, he says.

Outputs must be accurate, auditable, and secure, with users confident in results, says King from Taylor Rose.

Bhavisa Patel, Director of Legal Technology and Business Services at Eversheds Sutherland, says AI provides a great starting point but:

The ability to verify matters more than raw accuracy, says Birch from Redkite Solicitors.

"Lawyers trust tools that can show where outputs come from, such as sources, citations, document links, and confidence markers."

At the same time, clients are pressing for more transparency around how AI is used as well as more controls and verification. The top strategic priority for in-house counsel in 2026 is setting rules and safeguards for AI tools and products. Of all in-house corporate counsel, 39% said this is a priority, while 38% said it is something they're already doing.

Firms need to consider the impact an AI investment has on client relationships, regulatory compliance and their reputations.

“Trustworthy AI means putting in place systems, services and platforms that lawyers can rely on to support their professional judgement, without risking client interests, regulatory compliance or the firm’s reputation.”

Full transparency around how AI works is essential to building trust, says Lang from Fieldfisher.

"Lawyers need to understand how the AI arrived at a conclusion. Security, auditability and the extent to which humans remain in control all shape whether a system feels trustworthy.”

He adds that client data is now one of the most sensitive points in the conversation.

“Visibility around the use of client data, especially when it comes to training AI models, is paramount. Clients want to understand how firms will use their data today, but also how it may be used in the future.”

AI also needs to operate within a secure environment that safeguards client data and confidentiality, says Rendall from Taylor Wessing.

This forward-looking approach is becoming increasingly appealing to clients.

Capell from Gateley says: "Trust depends on robust safeguards: careful tool selection, disciplined workflow design, and a professional culture that reinforces the lawyer’s duty to review, challenge, and take ownership of AI assisted work."

Clients' interest in how firms use technology is a healthy and very welcome evolution, says Bazin from Lewis Silkin.

"The firms that can have this type of conversation calmly and early, so they stay on the front foot, will pull ahead of those who are still hoping nobody asks."

Santos at Bird & Bird shared a similar sentiment:

"That demand for confidence and transparency is pushing firms to tighten governance, sense-check tools, and be more open about trade-offs, not just benefits."

Trust will grow through greater transparency around how AI generates answers, but it also depends on effective governance, including clear policies, strong engagement and a well-managed rollout.

Earning trust through effective governance

Trust will come down to the credibility of the platform, the ease of access to certified sources, and familiarity of the workflow.

Across all respondents, 72% say they feel more confident using AI that’s grounded in legal sources. This rose to 79% for lawyers at large firms and 85% for in-house corporate lawyers.

When there are high technical risks, lawyers show a strong preference for legal-specific AI tools. This was evident across the most common use cases: large-scale document review for discovery and due diligence (42%), legal research (35%) and knowledge management (35%).

While mixed usage of legal and generic AI shows firms are experimenting, the figures indicate specialist legal platforms remain the dominant choice for core legal work.

Good AI governance starts with clear, visible ownership at board-level, says Taylor Rose's King.

"Essentially, this means moving beyond abstract principles and controls to tangible actions across standards for data quality, oversight of model selection and use cases, risk processes, and auditability to enable value at scale."

Trustworthy AI is AI that lawyers understand, can explain, and remain accountable for, says Santos from Bird & Bird.

"Lawyers trust tools that are transparent about data use and limitations, are tested in real legal workflows, and keep human judgement firmly in the lead."

In practice, governance creates trust as much as technology does, says Birch at Redkite Solicitors.

"Firms typically build confidence by starting with low risk activities such as internal drafting, brainstorming, and document summaries. This helps identify limitations and training needs before moving into medium risk work like first pass research or contract review support, and only later considering higher risk uses such as client advice."

Fieldfisher's Lang describes good AI governance as a careful balance between control and innovation.

That balance is difficult to maintain, he says, and governance cannot be treated as a one-off policy exercise.

“The scales will inevitably tip from one side to the other, but firms can reach equilibrium through continuous improvement, human intervention and transparency around what the AI does and how it achieves it.”

Legal AI tools are now everywhere, but not all AI is equal, and not all AI deserves your trust, says Stuart Greenhill, Senior Director of Segments at LexisNexis UK. 

"In legal work, confidence is not enough. Authority matters. Validation matters. Security matters. If you cannot stand behind the output, it is not legal AI. It is just AI".

With this shift, some firms have begun implementing more formal AI policies. John Craske, Chief Innovation & Knowledge Officer at CMS, describes an AI policy model that goes beyond experimentation:

What distinguishes this approach is that it treats AI as part of the organisation's operating philosophy. As Craske explains, it’s very important to identify the areas within workflows that humans must lead:

“Core human skills like emotional intelligence, critical thinking, balancing risks, getting the tone right, understanding what the client wants to achieve and why, negotiating, and getting deals done will be things that our people will play a central role in for years to come”.

Craske says good day-to-day AI use is pragmatic, well-governed, and value focused.

"It’s about using the right tools and resources for the job, understanding what the tools can and can’t do well, and working with them.”

Pairing the right AI tool with relevant human expertise will reshape what value looks like in the modern era.

Capell from Gateley says legal work is unique in that it is characterised by variation: "multiple lawyers often perform “the same” task in subtly different ways."

AI optimists sometimes underestimate how much effort is required to align processes, behaviours, and incentives before the technology, however clever, can be deployed successfully, he says.

"Scaling AI requires a broader skill set than technology alone: legal process design, operational leadership, training, and ongoing engagement are just as critical as model performance."

As AI becomes more capable and increasingly more agentic, good governance needs to be evolutionary rather than static, says Capell.

"We need governance that can scale. In particular, governance now needs to address three emerging challenges: how to supervise semi-autonomous workflows, how to manage and reuse end-user generated prompts and AI driven processes, and how to align proportionately with regulatory frameworks such as the EU AI Act."

We treat generative AI as a core professional competency rather than a peripheral tool, he adds. "That means supporting experimentation and learning, while also establishing shared standards."

Patel from Eversheds highlights the wider cultural dimension of this shift:

To put it bluntly, it's not just about the cleverness of an AI model or even the quality of the content. It's the system around it, the reliability of retrieval, evaluation of process and review agents that make technology trustworthy.

Introducing The Four Layer Trust Stack

Trust is not a fancy feature of AI. It is a system to develop and build. To demonstrate, we've designed the Four-Layer Trust Stack for CTOs to aid in embedding trust when implementing legal AI workflows.

Infrastructure Trust.  This is the area where CTOs should feel the most comfortable navigating. It's the security and privacy of the technology, as well as its architecture, access controls, risk mitigations and data auditability.

Technical Trust.  This comes down to the quality and depth of legal AI tools. What content is it grounded on? How easy is it to access citations that add value? Where does the underlying content come from? This needs clear audit trails that enable lawyers to see how an answer has been reached, appropriate evaluation, testing, proof of concept and tracking of results.

Oliver Bethell, Chief Technology Officer at Travers Smith, says CTOs need to think about trust across both the platform layer and the model layer.

“There are two angles to trustworthy AI: trusting that data is secure, and trusting that models produce reliable answers.”

That means enterprise-grade controls around data privacy, information security and retention, alongside a clear understanding of how model capability is changing over time.

Workflow Trust.  This layer looks at where AI fits into the way lawyers actually work. The review process, or legal research workflow, or drafting a document. The metric that matters most here is not speed of answer. Time-to-safe-answer is the real KPI that matters.

Human Trust.  This is the most pressing layer, and one that most firms are still stuck on. This includes change management, incentives, training, confidence, behaviour shifts and creating a culture where AI use is governed and used properly, not improvised.