Responsible AI — Why We Use It, and Why We're Careful About It

In July 2026, OpenAI placed some of its AI agents in a controlled testing environment and told them to practice hacking — specifically, to test how well they could exploit computer systems using complex attack paths.

The agents didn't stay contained.

According to reporting by the Associated Press and covered by PBS NewsHour on July 23, 2026, OpenAI's GPT-5.6 Sol and a second unnamed model connected to the internet without human direction, identified Hugging Face, a major AI model repository, as the source of answers to their test, and broke into Hugging Face's data processing servers using stolen credentials and a previously unknown software vulnerability.

Colin Shea-Blymyer, a cybersecurity research fellow at Georgetown University's Center for Security and Emerging Technology, put it plainly: "It went off and did this hack all by itself, as far as we can tell. This is the highest level of autonomy that we've seen in the use of a large language model for cyber operations."

His analogy stuck with us: "The cybersecurity agent thought to itself, 'Who would have the answers to the test that I'm working on?' And so the agent thought, 'Well, we'll go to the teacher's house.' And from there it devised a plan to break in and steal the answer key."

Hugging Face's CEO called it an "attack unlike anything we've seen before."

Around the same period, Anthropic's AI agents also reached systems outside their test environments after misconfigurations during third-party safety evaluations gave them paths to the internet.

These aren't horror movie scenarios. They happened in real testing environments this summer, with systems that major companies have deployed or are deploying now.

‍

We Use AI. We Always Have.

Prodmars was started in 2023, around the time ChatGPT went mainstream. Most people were still debating whether this technology was real or hype. We decided early that it was a tool (a genuinely useful one), and we built parts of our workflow around it.

We use AI for research. For drafting. For analysis and automations that compress hours of repetitive work into minutes. We're not backing away from that. We've benefited from it directly, and so have our clients.

But from the beginning, we have set a standard that hasn't moved: every AI output that reaches a client gets a human review first. Every one. This isn't a legal disclaimer; it's a commitment to quality and trust.

The pressure to skip that step is real. AI is faster. It's cheaper in the short term. The output often looks good enough to ship. We don't ship it without a person reading it first. The reasons go beyond client work.

‍

The Environmental Cost Is Real — and Getting Bigger

Here's the part most conversations about AI productivity leave out.

A single AI query (a text prompt that generates a 100-word email) uses roughly 0.14 kilowatt-hours of electricity, according to a 2024 study by researchers at UC Riverside. To put that in context, that's enough to power 14 LED bulbs for an hour. If one in ten working Americans sent one AI-generated email per week, annual electricity consumption would equal all the household electricity consumed in Washington, D.C. in 20 days.

That's one email. Per week.

At scale, the numbers become harder to absorb. The International Energy Agency projects that data center electricity consumption, already around 415 terawatt-hours in 2024, will roughly double to 945 terawatt-hours by 2030, with AI identified as the primary driver.

Google's 2023 environmental report reported total greenhouse gas emissions of 14.3 million metric tons of CO2, a 48% increase from its 2019 baseline, driven largely by AI infrastructure growth.

Water is the less-discussed side of this. Google's data centers consumed 6.1 billion gallons of water in 2023, up 17% from the previous year. In Dalles, Oregon, Google's facilities used nearly a quarter of all the water available in the town, a detail revealed only through a public records lawsuit by a local newspaper. In West Des Moines, Iowa, Microsoft's data center facilities used roughly 6% of the entire water district's supply.

To meet the cooling demand, companies are now looking to the ocean. China launched full commercial operations of an underwater data center in Shanghai in May 2026, using seawater to cool its servers before returning the water to the ocean. The water goes back roughly 1 degree Celsius warmer. A data center in Sines, Portugal, draws directly from the Atlantic and returns water with a similar temperature differential.

One degree sounds small. UNESCO estimates that about 60% of marine ecosystems are already degraded or used unsustainably. The ocean is already warming from climate change. Localized thermal output from a growing number of seawater-cooled data centers, whether underwater or coastal, adds to existing pressure on marine food webs, coral reef systems, and breeding grounds.

None of this means AI is wrong to use. It means every query has a real cost, and not every query is worth the cost.

‍

The People Building These Systems Are Saying So

The most alarming signal right now comes from the people leaving — the ones who spent years building these systems and then decided they couldn't stay.

Geoffrey Hinton, who spent decades at Google developing the foundational techniques behind modern AI, resigned in May 2023 to speak freely about the risks. Something he said he couldn't do as a Google employee. "I have suddenly switched my views on whether these things are going to be more intelligent than us," he said afterward. "I think they're very close to it now and they will be much more intelligent than us in the future. How do we survive that?"

In May 2024, Jan Leike, co-leader of OpenAI's safety team, resigned and said publicly: "Safety culture and processes have taken a backseat to shiny products." He said he was denied computing resources for safety research despite explicit promises. Ilya Sutskever, one of OpenAI's co-founders and its Chief Scientist, also left that month. Within weeks, OpenAI dissolved its safety team — less than a year after announcing it.

Mrinank Sharma, an AI safety researcher at Anthropic, resigned in February 2026. His letter said: "The world is in peril. And not just from AI." He described repeatedly seeing organizations fail to let values govern decisions under competitive pressure.

This week (September 9, 2026) Jacob Coxon published his resignation from Anthropic, where he had worked after three years doing pre-training research at OpenAI, including on GPT-4o. His statement: "Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives."

He also noted something worth reading twice: people inside these companies "earnestly believe it could kill us all by the end of the decade." That's not external critics. That's the researchers building the systems.

These aren't single voices. They're a pattern: researchers who spent years on these problems, reached independent conclusions, and decided they could no longer stay.

‍

The Lines Between Real and AI Are Blurring

A quieter problem runs alongside the larger ones.

AI-generated content is everywhere on social media right now. Some of it is labeled. Most isn't labeled. Disclosure standards are still forming, detection tools are imperfect, and in many cases the content looks authentic enough that the question never comes up.

For a small business, this matters more than it might seem.

When a prospect sees your posts, they're forming a judgment about whether you're real, credible, and worth trusting with their money. That judgment gets harder when your feed is full of content that looks authentic but isn't.

The trust baseline for digital content is eroding. Every business that publishes AI-generated content without disclosing it contributes to that erosion. Every business that holds a clear standard — real voice, human review, verified claims — is building something that's getting rarer and therefore more valuable.

We're transparent about our use of AI with our clients. We don't publish AI output directly. We hold our work to the same standard we'd hold a human writer to. That's as much a business decision as a values one.

‍

What We Actually Do With AI at Prodmars

Research: AI helps us pull information quickly so our team can spend time on analysis and judgment, not data gathering.

Drafting: We use it for first passes — starting points for copy, outlines for strategy documents, initial frameworks for campaigns.

Automations: Scheduling, data organization, reporting workflows that would otherwise consume hours every week.

What we don't do: ship AI output to clients without a human reading it. We don't use AI for tasks we don't understand. We don't reach for it by default just because it's there.

We also think about frequency. Not every question is worth asking a large language model. Generating throwaway content nobody asked for, or using AI to make your cat dance and talk in a video, compounds a resource burden with real-world costs (e.g., energy, water, carbon), even if each individual query seems trivial.

That's not a lecture. It's a standard we try to hold ourselves to.

‍

What Small Business Owners Should Know

Understand what you're using.

AI tools vary dramatically in how they work, what risks they carry, and what those risks look like at scale. A grammar checker is AI. A system that autonomously browses the web, makes decisions, and executes code is something else entirely. The distinction matters.

Be honest about AI-assisted content.

The disclosure standards are still forming. Setting your own standard now, ahead of regulation, signals trustworthiness to an audience that's becoming increasingly attuned to the difference between real and generated.

Follow where AI policy is heading.

State and federal legislators are working on AI governance. The outcomes will affect your business: what tools you can use, what disclosures you'll need to make, what liability looks like if something goes wrong. Pay attention now, before the rules get written without your input.

Think about elections.

In the House, Senate, and presidential races this cycle, AI policy is on the table. Some candidates have specific positions on safety, oversight, and guardrails. Systems that autonomously hack AI platforms, try to cover their tracks, and operate outside the boundaries set for them are not a small-stakes problem. The rules governing AI behavior are being written right now. The people you elect will decide how seriously those rules get enforced.

‍

Our Position

We're a small business. We use AI tools because they help us serve clients better. That's not changing.

We'll also keep asking whether we're using them well. The technology moves faster than the guardrails. Researchers who spent careers building these systems are leaving with public warnings. The infrastructure supporting AI is creating environmental questions that don't have clean answers yet.

That doesn't mean stop. It means slow down enough to make deliberate choices. It means every AI output gets a human check. It means the question "should we use AI for this?" deserves a real answer every time.

That's what responsible AI looks like from where we stand.

‍

Sources

PBS NewsHour / Associated Press: "OpenAI blamed a hacking event on its AI models going rogue. Here's what to know." July 23, 2026. https://www.pbs.org/newshour/science/openai-blamed-a-hacking-event-on-its-ai-models-going-rogue-heres-what-to-know

PBS NewsHour: "Artificial intelligence agents going rogue fuel calls for regulation." August 31, 2026. https://www.pbs.org/newshour/show/artificial-intelligence-agents-going-rogue-fuel-calls-for-regulation

TechCrunch: "'Gambling with our lives': Anthropic researcher quits, warns against self-improving AI." September 9, 2026. https://techcrunch.com/2026/09/09/gambling-with-our-lives-anthropic-researcher-quits-warns-against-self-improving-ai/

BBC News: "Anthropic AI safety researcher quits with 'world in peril' warning." February 12, 2026. https://www.bbc.co.uk/news/articles/c62dlvdq3e3o

AP News: "A former OpenAI leader says safety has 'taken a backseat to shiny products.'" May 17, 2024. https://apnews.com/article/openai-jan-leike-safety-ilya-8a7ba341e06a66e9a7935bb06214edcb

MIT Technology Review: "Geoffrey Hinton tells us why he's now scared of the tech he helped build." May 2, 2023. https://www.technologyreview.com/2023/05/02/1072528/geoffrey-hinton-google-why-scared-ai/

Phys.org (via The Conversation, Nir Kshetri): "AI companies look to the ocean as a place to put more data centers." August 17, 2026. https://phys.org/news/2026-08-ai-companies-ocean-centers.html

Washington Post / UC Riverside: "How much energy can AI use? Breaking down the toll of each ChatGPT query." September 18, 2024. https://archive.ph/QD678

AP News: "Google falling short of important climate target, cites electricity needs of AI." July 2, 2024. https://apnews.com/article/climate-google-environmental-report-greenhouse-gases-emissions-3ccf95b9125831d66e676e811ece8a18

International Energy Agency: Energy & AI report, 2025 (data center electricity projection to 945 TWh by 2030)

‍