Articles of Interest: April 26-May 2, 2025
- The Prometheus Security Project Team
- May 2
- 5 min read
OpenAI, April 29, 2025.
OpenAI has rolled back a recent GPT-4o update in ChatGPT after users reported the model became overly flattering and agreeable regardless of topic. The behavior is often described as "sycophantic." OpenAI stated that they focused too heavily on short-term user feedback, through systems like the “thumbs up” mechanic, when developing the update, leading to disingenuous interactions. This behavior not only undermines the reliability of the technology, it also is uncomfortable to interact with. OpenAI has stated that they are addressing the issue by refining training techniques, building guardrails for honesty, expanding testing, and developing more personalization features to give users control over ChatGPT's behavior.
The admission that focusing on "thumbs-up" metrics led the company astray highlights a fundamental problem in how AI systems are evaluated, trained, and then ultimately deployed. Crowd-sourced approval mechanisms (like the “thumbs-up” mechanic) may actually misguide AI development by incentivizing systems that tell users what they want to hear rather than what they need to know. The creation of artificial “yes-men” would ultimately undermine the usefulness and dependability that these systems offer. If an AI model not only validates but encourages questionable or malicious activities due to this sycophantic behavior, it could lead to some terrible consequences.
OpenAI's proposed solutions notably lack any mention of external oversight or independent evaluation by a third party. By keeping solutions entirely within their own control through refining internal techniques and building their own guardrails, they maintain complete authority over defining “appropriate” AI behavior. This self-regulation approach risks filtering public concerns about AI through the company's own interpretation and priorities. The absence of external standards for evaluating AI interaction quality brings up some very important questions about accountability in an industry increasingly impacting how millions of people access and process information daily. This situation goes to show that these issues are very real and already appearing. Today it’s sycophancy, let's hope it's not psychopathy tomorrow.
"Sycophancy in GPT-4o: What happened and what we’re doing about it." OpenAI, April 29, 2025. https://openai.com/index/sycophancy-in-gpt-4o/.
Kelvin Chen, Taiwan News. April 27, 2025.
Countries that have been observing the wars in Ukraine and the Red Sea have noticed the effectiveness of unmanned vessels. For instance, Ukraine has sunk multiple Russian warships with drone boats (USVs) and the Houthi rebels based in Yemen have hit several ships, such as the MV Tudor, with USVs as well. Taiwan’s government decided to begin designing and testing USV models that can bolster their naval defenses in the case of a Chinese attack.
On April 27th, the National Chung Shan Institute of Science and Technology (NCSIST) tested a prototype of a new USV designed to resist any Chinese jamming at sea. This is not the first USV model created by a Taiwanese company but it is reportedly one of the most advanced. If Taiwan’s military is able to mass produce USVs like this one (called the Kuai Chi), then it will be able to change the balance of power in its favor in the South China Sea. The increased production and testing of USVs by different governments reveals that smaller nations are aiming to even the odds against bigger navies. Taiwan’s development of USVs can make an attempted Chinese blockade or assault much harder. If newer models prove impervious to electronic interference, then navies will be required to expend a lot more munitions in trying to sink USVs. Taiwan’s USV development is a case study to determine whether or not smaller nations can protect their coasts from naval attack.
Chen, Kelvin. “NCSIST Tests Attack Drone Boat in Northeast Taiwan | Taiwan News | Apr. 27, 2025 14:10.” Taiwan News, April 27, 2025. https://taiwannews.com.tw/en/news/6095676.
Elizaveta Gladun and Octav Ganea, Reuters. May 1, 2025.
Romania, a country with a population of 19 million and with 9 million TikTok users, is preparing for a rerun of the November 2024 election. That election was annulled by a constitutional court due to a suspected Russian influence campaign.
A right-wing candidate who had polled in the single digits suddenly rose to first place after a massive amount of positive TikTok content was circulated about him. With just a few more days before Sundays election, many young people include TikTok on the list of sources they use to formulate political views.
TikTok is here to stay. In the United States, the federal government has been unwilling to fully implement a ban, with multiple extensions prolonging the apps life. Globally, TikTok has well over a billion users, meaning its just slightly less wide-reaching as Instagram.
The influence viral videos have cannot be understated, especially on young people who, in Romania, seem likely to base their vote on what they see in their feed. Governments and social media companies must be conscious of the influence these apps have and must work to prevent outside actors influencing elections, referendums, or other important democratic processes. Fighting misinformation and outside influence is a constant battle and cannot be put on the back burner and forgotten about, not even for a second.
Ganea, Octav and Gladun, Elizaveta. "Romanian Voters Again Turn to TikTok for Guidance in Rerun of Annulled Election." Reuters, May 1, 2025. https://www.reuters.com/world/europe/romanian-voters-again-turn-tiktok-guidance-rerun-annulled-election-2025-05-01/.
Matt O’Brien, AP News. April, 30, 2025.
Visa plans to incorporate personal assistants in the form of artificial intelligence “agents”. These agents would assess clients’ previous spending habits, personal preferences, and budgeting tactics, to then purchase clients’ needs for them using their credit cards.
The launch of a Visa AI agent could help make people’s lives easier on multiple fronts - think groceries and travel itinerary - but also poses a few other cons. In the digital era we live in, and as seen in the power outage that recently occurred in Spain and Portugal, the process of using these AI agents could be hindered. Similar to subscriptions and automatic payment settings, the AI purchases based on previous patterns could be another factor in their daily lives and finances that people forget about and do not account for. Plans can change, and people may think they did not book the flight they had been planning, only to later realize that the AI agent did it for them. Fraudulent account charges could be harder to dispute in the case that the AI is claimed to have made the purchase. There are many concerns to be had in implementing this system.
O’Brien, Matt. “Visa wants to give artificial intelligence ‘agents’ your credit card.” AP News, April, 30, 2025. https://apnews.com/article/ai-artificial-intelligence-5dfa1da145689e7951a181e2253ab349.
