| | Making sense of… | | An Everyday Guide to Using AI Safely | | People have mixed feelings about AI. We see the passionate backlash in different parts of the world and genuine excitement for many people, including ourselves. The Internet and traditional media seem to be generally replete with practical advice, and it mostly comes from AI vendors, influencer content that promotes products or services, or material that is very cautionary but doesn’t tell you what to do differently. There’s also some purely anti-AI advice that bluntly states not to use the latest generative AI tools. We couldn’t find a guide that was rigorous, vendor-neutral, backed by evidence, and written for the ordinary person trying to build healthy and safe habits. This is why we wrote one. We wrote The Everyday Guide to Using AI Safely. The document is organised around the premise that you have more agency in how you use AI than most advice acknowledges. We cover things like understanding where AI tends to be reliable, where it routinely fails, and what habits reduce your exposure. We’ve curated the further reading sources ourselves and stand behind them. This is version one, and we are genuinely seeking reader feedback to further improve it. | | Read the full guide → | |
|
| | On the radar | | 1. In his recent speech at The University of Sydney, Australian Prime Minister Albanese has announced a set of Australian Standards for AI and a new Office of AI sitting within the Department of the Prime Minister and Cabinet. Combined with five consumer safety priorities (covering duty of care, privacy, workplace safety, consumer protections, and automated decision-making), this is a positive domestic policy move. We’re pleased to see a recognition that AI does not only belong to industry policy (where meaningful work has been done already over the years) but to societal policy, and has rightly been elevated to that level. It is worth noting that work on the consumer safety areas was already underway - in some cases for years - but it is reassuring to see the visible recommitment. ABC News | | 2. Word is spreading among government AI circles globally about Alberta’s approach documented in their AI Velocity White Papers. Some government staff are looking to it for inspiration. The papers document the entire journey of the Alberta government maintaining and modernising a software codebase that surpassed the point where it can be maintained solely by human work. They describe in great detail the process they’re following, what is working, learnings from the codebase, governance approaches, how they go about measuring and reporting efficiencies, cost, and time savings. | | Trendline. In our previous Gradient Brief, we identified the ‘task horizon’ - the duration over which AI agents can operate reliably - as a critical metric to monitor. This window is widening rapidly, granting models the capacity to execute increasingly complex, multi-step cyber operations. What was a theoretical risk has now manifested: Hugging Face, which provides infrastructure for hosting AI models, recently contained a security breach stemming from a fully autonomous AI agent. OpenAI soon confirmed the attack originated from an internal research sandbox during live evaluations. As both firms collaborate on a forensic review, this incident serves as a stark warning for what capabilities are out there, and the analysis by Hugging Face raises important learnings for defenders. |
| |
|
| | Question of the month | | How should senior leaders think about the risks of using open versus proprietary AI models? | | Advocates for open-weight models point to transparency, auditability, and sovereign infrastructure. For Hugging Face, utilising a self-hosted model provided two key advantages during the incident mentioned above: it allowed them to bypass safety guardrails backed into proprietary models that might have otherwise blocked their response, and it enabled them to contain the breach, ensuring that sensitive credentials and data remained within their secure environment. Advocates for proprietary models point to vendor accountability, faster security patching, contractual recourse, and the practical reality that most organisations lack the in-house talent or bandwidth to secure an open-weight deployment themselves. Are you debating these issues in your organisation? We read every answer. | | Reply to share your thinking → | |
|
Media enquiries For all enquiries (including media, speaker, education, advisory and research), please email info@gradientinstitute.org and address it to Sarah. |
|
|
