Telescope launches Guardrails, powered by GPT4-o, for ethical AI in finance

By Sophie Moore on

Today, we’re excited to announce the launch of Guardrails, our AI enabled product for compliance and ethical content screening in financial applications. Guardrails is powered by GPT4-o, the groundbreaking new model from OpenAI that was just unveiled last night.

Safety at its core

Guardrails operates within the confines of your compliance and ethical frameworks to screen for potentially problematic inputs in finance. Our system screens prompts for sensitive content, such as hate speech, racial issues, misinformation, political biases, illegal activities, and violence. It also safeguards against inadvertent financial advice with a sophisticated advice-filtering system that can handle diverse user inputs.

We understand that regulatory requirements vary by jurisdiction. That’s why we work closely with our clients to tailor Guardrails to meet the specific needs of their area. Our team has invested hundreds of hours in testing, engineering, and building observability pipelines to ensure Guardrails upholds the highest standards of safety and compliance.

”Building a safe AI system for finance is an immense challenge that frankly, we initially underestimated,” says Luc Pettett, Telescope’s founder. “It’s not enough to just filter out blatantly inappropriate content. We have to consider edge cases and gray areas. We don’t want users searching for how to invest in nuclear warfare or the viability of penguin meat. And personalized financial advice is completely off the table. Getting this right has required extensive testing against thousands of prompts, including input from our private beta users.”

Benchmarking Examples Benchmarking Telescope Guardrails.

Defining acceptable boundaries in finance is a complex undertaking. “What’s considered appropriate can vary widely depending on the context and the user,” explains Megan Fawdon, Telescope’s Operational Manager. “We benchmark against multiple models and manually screen each prompt, iterating on our own chain of thought approach. It’s a continuous process of refinement to uphold both the letter and the spirit of ethical, compliant AI in finance.”

Guardrails sits on top of our flagship product Ripple, which is already in use by several large retail stock brokers and publishers, with public product launches imminent. Ripple allows investors to discover stock ideas from natural language prompts, and Guardrails ensures this powerful capability remains within appropriate boundaries.

Deep assessment capabilities

Guardrails goes beyond simple keyword filtering to provide deep, contextual screening of longer-form text. Our advanced AI models can analyze entire articles, reports, or user-generated content to identify problematic passages that may be misleading, biased, or non-compliant. But Guardrails doesn’t just flag issues - it provides intelligent recommendations on how to address them. Whether it’s suggesting alternative phrasing, providing additional context, or recommending the removal of certain sections, Guardrails offers actionable insights to help our clients maintain the highest standards of content integrity.

Benchmarking Example: Advice Verifing a news article for potentially problematic financial advice.

Powered by GPT4-o

The launch of Guardrails coincides with the release of GPT4-o, the latest breakthrough from OpenAI. GPT4-o represents a major leap forward in natural human-computer interaction, accepting any combination of text, audio, and image inputs and generating corresponding outputs. It matches GPT-4 Turbo’s performance on English text and code while significantly improving non-English language understanding. Importantly for a real-time application like Guardrails, GPT4-o can respond to audio inputs in as little as 232 milliseconds, on par with human response times in conversation.

Safety and transparency are core priorities with GPT4-o. The model has undergone rigorous testing and red teaming to identify and mitigate risks, especially those unique to the new audio modalities. OpenAI is taking a phased approach to rolling out GPT4-o’s capabilities, starting with text and image inputs and text outputs. Audio outputs will initially be limited to preset voices in line with OpenAI’s safety policies.

While tremendously powerful, GPT4-o does have limitations that OpenAI is proactively addressing. The model can occasionally produce incorrect information or make reasoning errors. It may also exhibit biases and generate unsafe or biased content, although extensive work has gone into mitigating these issues. OpenAI emphasizes that GPT4-o, like all AI systems today, should be used as a complement to human intelligence rather than a replacement.

The future of AI in finance

At Telescope, we believe AI has immense potential to make finance more accessible, personalized, and engaging for investors. But realizing that potential requires an unwavering commitment to safety, compliance, and transparency. With the launch of Guardrails powered by GPT4-o, we’re taking a major step forward in our mission to inspire the future of investing through responsible AI innovation.