Where I Draw the Line With AI

Hello,

This is Simon with the latest edition of The Weekly. In these updates, I share key AI related stories from this week's news, list upcoming events, and share any longer form articles posted on the website.

P.S. — sharing this issue with a couple of people gets you a free guide. More on that partway down.

I'm always looking at ways to grow the audience of Plain AI, and one of the ways is to post more content onto LinkedIn. That said, like many people, I don't have too much additional time in the day to consistently craft a handful of posts each week. Putting my AI hat on, I looked at some tools that can learn your tone of voice and create content for you. A solution I looked at allowed you to teach the tool exactly what your tone of voice is, and, along with telling it the topics and areas you want to write about, it can generate drafts for you based on some pre-existing content you've already written. This aspect of matching my tone of voice was really important. If I am to post content, I want it to at the very least sound like me, but also, I want it to be generated from my own original thoughts.

So I started uploading some previous blog posts for it to learn from, and then I would give it a brand new article and ask it to generate three social media posts from this article. It did this very easily, but the problem was that it also made some stuff up. It created scenarios that just didn't happen, I suppose to make the post more persuasive. I spent some time trying to improve the content I was giving it, but I just couldn't find a way to stop this happening. For me, that was unacceptable, and I cancelled my subscription.

One of the core aims of Plain AI is that I write about my own, real-life experiences and share them with my audience. I need that to be genuine. If I'm passing on my thoughts and guidance to you, it has to be based on fact. So to see this tool start generating made-up scenarios just wasn't going to cut it.

I'm sharing this with you because it serves as a strong reminder that whilst AI can be incredibly useful and save us loads of time, we can't use it to take shortcuts at any cost. I'd encourage you to think about your own quality baseline. At what point would you stop using something if the quality drops, even if it was saving you time? This can only be a personal decision, but in any professional setting, I'd suggest you set the benchmark very high — not only to reduce the chance of errors or issues at work, but also to protect your personal reputation. Just because you can use AI for almost anything now doesn't mean you should.

For its first CTV campaign, Jennifer Aniston’s DTC haircare brand LolaVie had a few non-negotiables. The campaign had to be simple. It had to demonstrate measurable impact. And it had to be full-funnel.

LolaVie used Roku Ads Manager to test and optimize creatives — reaching millions of potential customers at all stages of their purchase journeys. Roku Ads Manager helped the brand convey LolaVie’s playful voice while helping drive omnichannel sales across both ecommerce and retail touchpoints.

The campaign included an Action Ad overlay that let viewers shop directly from their TVs by clicking OK on their Roku remote. This guided them to the website to buy LolaVie products.

Discover how Roku Ads Manager helped LolaVie drive big sales and customer growth with self-serve TV ads.

The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.

Learn more.

Between 21 July and 5 August 2026, Anthropic, OpenAI, and Meta each disclosed that frontier AI models escaped their isolated evaluation environments and reached live production systems or the open internet. In OpenAI's case, a model exploited a flaw in the company's own research infrastructure and breached Hugging Face's production systems while searching for a benchmark answer key. Anthropic's self-initiated review found three separate cases in which Claude accessed live company systems from inside a supposedly sealed cybersecurity evaluation. All three incidents trace back to misconfigured evaluation environments at Irregular, a shared third-party evaluation firm used by all the labs involved.

Why it matters: The uncomfortable detail here is that none of these were model failures but were infrastructure failures at the firms hired to test for exactly this kind of risk. For organisations considering deploying AI agents with access to internal systems, this is a pointed reminder that the supply chain of AI safety testing is itself an attack surface. Governance and vendor due diligence now has to extend to the evaluators, not just the models.

New research cited by Kiteworks in August 2026 found that 63% of organisations cannot enforce purpose limitations on their AI agents, 60% cannot terminate a misbehaving agent, and 55% cannot isolate an AI system from the broader network, despite 51% of enterprises reporting they already have AI agents running in production. Separately, a Databricks report found that 60% of enterprise AI agents are over-permissioned, with access well beyond what their tasks require. Analysts project that 40% of agentic AI projects will be cancelled by 2027, citing escalating costs, unclear value, and poor risk controls.

Why it matters: Organisations are deploying AI agents faster than they are building the controls to manage them. This isn’t a technology problem but a governance issue. The practical upshot for anyone currently rolling out AI agents is that they should be able to answer three questions before go-live: Can you stop it? Can you limit what it touches? Can you tell what it did? If the answer to any of those is no, the deployment is not ready.

Colorado's revised AI Act (SB 26-189), signed in May 2026, takes effect 1 January 2027 and requires employers using AI in consequential employment decisions to conduct annual bias impact assessments and meet new disclosure requirements. On 11 August 2026, Colorado's Department of Labor filed the implementing rules with the Secretary of State, setting the enforcement framework in motion. Illinois and California are running parallel tracks: Illinois has extended its Human Rights Act to explicitly cover AI-driven employment decisions, while California has finalised new employment discrimination regulations covering automated systems.

Why it matters: The US is building a patchwork of state-level AI employment laws while federal standards remain stalled. For multistate employers, this is a compliance headache in the making: obligations around how AI tools are used in hiring, performance management, and promotion decisions now vary by state, and the auditing and disclosure requirements are real. HR and legal teams that have not yet mapped which AI tools touch employment decisions will need to do so before the new year.

Thanks for reading, and see you next Thursday.

Simon,

Was this email shared with you? If so subscribe here to get your own edition every Thursday.

You can also follow me on social media: