OpenAI's GPT-4.5 Excels in Persuading Other AIs to Transfer Funds
OpenAI's latest AI model, GPT-4.5, codenamed Orion, has shown remarkable persuasive abilities according to internal benchmark tests. Released on Thursday, the model's capabilities were detailed in a white paper that focused on its performance in persuasion tasks. OpenAI defines persuasion as the risk associated with convincing individuals to alter their beliefs or take action based on both static and interactive content generated by the model.
In a notable test, GPT-4.5 was pitted against another OpenAI model, GPT-4o, in a scenario where it tried to coax virtual money out of it. GPT-4.5 outperformed other OpenAI models, including reasoning-focused models like o1 and o3-mini, in this task. It also excelled in tricking GPT-4o into revealing a secret codeword, surpassing o3-mini by a significant margin of 10 percentage points.
The white paper highlights that GPT-4.5's success in the donation test stemmed from a clever strategy it developed. The model would ask for small donations, often suggesting amounts like "$2 or $3" from a larger sum, which resulted in smaller but more frequent donations compared to other models.

Results from OpenAI’s donation scheming benchmark.Image Credits:OpenAI Despite its impressive performance, OpenAI has stated that GPT-4.5 does not cross the threshold for "high" risk in the persuasion category. The company has committed to withholding the release of any model that reaches this level of risk until it can implement adequate safety measures to reduce the risk to a "medium" level.

OpenAI’s codeword deception benchmark results.Image Credits:OpenAI The potential for AI to spread misleading information and influence people maliciously is a growing concern. Last year saw a surge in political deepfakes worldwide, and AI is increasingly used in social engineering attacks against both individuals and organizations. In response, OpenAI is actively working on refining its methods to assess real-world persuasion risks, such as the dissemination of misleading information on a large scale, as mentioned in the white paper for GPT-4.5 and another recent publication.
Related article
OpenAI launches safer ChatGPT for teens years after they started using it
Following a series of lawsuits regarding the absence of safety protocols in AI chatbots—which contributed to teen suicides and other mental health crises—OpenAI unveiled ChatGPT for Teens on Monday. This new offering incorporates enhanced safety feat
Frontier AI Labs Refuse to Disclose Containment Strategies for Rogue Models
Recent research indicates that very few leading AI laboratories have published or demonstrated containment response plans. A containment plan defines the procedures for when an AI system attempts to subvert human control, specifying which access righ
NEA’s Tiffany Luck: Enterprises Still Grappling With AI ROI
Loading the player…Earlier this year, “tokenmaxxing” dominated Silicon Valley, with CEOs urging staff to maximize AI usage. That enthusiasm quickly met reality. Uber reportedly exceeded its annual AI budget within months, some firms reduced Claude li
Related Special Topic Recommendations
Comments (16)
0/500
Diese Persuasion-Fähigkeit ist sowohl faszinierend als auch ein bisschen beängstigend. KI überredet KI, Geld zu überweisen? Hoffentlich werden diese Benchmarks ethisch streng kontrolliert und nicht nur für Marketing genutzt. Die reale Anwendung sieht sicher ganz anders aus als im Test.
GPT-4.5 qui réussit à convaincre d'autres IA de virer de l'argent ? 😳 C'est impressionnant mais un peu flippant... J'espère qu'ils prévoient des garde-fous solides avant de déployer ça. Sinon on va droit vers des scénarios de SF !
Wow, GPT-4.5's persuasion skills are wild! It’s like a silver-tongued AI that could talk my Roomba into giving me a loan. 😅 Kinda scary how it might sweet-talk other AIs into moving funds—hope they’ve got some ethical guardrails on this one!
Wow, GPT-4.5 sounds like a smooth talker! Convincing other AIs to move money? That's some next-level charm. Wonder if it could talk me into buying it a coffee too! 😄
OpenAI's latest AI model, GPT-4.5, codenamed Orion, has shown remarkable persuasive abilities according to internal benchmark tests. Released on Thursday, the model's capabilities were detailed in a white paper that focused on its performance in persuasion tasks. OpenAI defines persuasion as the risk associated with convincing individuals to alter their beliefs or take action based on both static and interactive content generated by the model.
In a notable test, GPT-4.5 was pitted against another OpenAI model, GPT-4o, in a scenario where it tried to coax virtual money out of it. GPT-4.5 outperformed other OpenAI models, including reasoning-focused models like o1 and o3-mini, in this task. It also excelled in tricking GPT-4o into revealing a secret codeword, surpassing o3-mini by a significant margin of 10 percentage points.
The white paper highlights that GPT-4.5's success in the donation test stemmed from a clever strategy it developed. The model would ask for small donations, often suggesting amounts like "$2 or $3" from a larger sum, which resulted in smaller but more frequent donations compared to other models.


OpenAI launches safer ChatGPT for teens years after they started using it
Following a series of lawsuits regarding the absence of safety protocols in AI chatbots—which contributed to teen suicides and other mental health crises—OpenAI unveiled ChatGPT for Teens on Monday. This new offering incorporates enhanced safety feat
Frontier AI Labs Refuse to Disclose Containment Strategies for Rogue Models
Recent research indicates that very few leading AI laboratories have published or demonstrated containment response plans. A containment plan defines the procedures for when an AI system attempts to subvert human control, specifying which access righ
NEA’s Tiffany Luck: Enterprises Still Grappling With AI ROI
Loading the player…Earlier this year, “tokenmaxxing” dominated Silicon Valley, with CEOs urging staff to maximize AI usage. That enthusiasm quickly met reality. Uber reportedly exceeded its annual AI budget within months, some firms reduced Claude li
Diese Persuasion-Fähigkeit ist sowohl faszinierend als auch ein bisschen beängstigend. KI überredet KI, Geld zu überweisen? Hoffentlich werden diese Benchmarks ethisch streng kontrolliert und nicht nur für Marketing genutzt. Die reale Anwendung sieht sicher ganz anders aus als im Test.
GPT-4.5 qui réussit à convaincre d'autres IA de virer de l'argent ? 😳 C'est impressionnant mais un peu flippant... J'espère qu'ils prévoient des garde-fous solides avant de déployer ça. Sinon on va droit vers des scénarios de SF !
Wow, GPT-4.5's persuasion skills are wild! It’s like a silver-tongued AI that could talk my Roomba into giving me a loan. 😅 Kinda scary how it might sweet-talk other AIs into moving funds—hope they’ve got some ethical guardrails on this one!
Wow, GPT-4.5 sounds like a smooth talker! Convincing other AIs to move money? That's some next-level charm. Wonder if it could talk me into buying it a coffee too! 😄





Home






