pytorch/examples

Pytorch is insufficiently opinionated

Opened this issue · 0 comments

Deleted user commented

🐛 Describe the bug

Context

Machine learning models can be trained on secret, synthetic, or biased data to create seemingly authoritative probability estimates used for abusive purposes in legal contexts. In Jessica Logan's case, her 911 call was used as "evidence" (when interpreted by a poorly trained and overly confident detective) that she had killed her baby.

As an example, an affiliate of Tracy Harpster [1] currently conspires to create an AI startup (Deceptio AI) to launder voodoo (via hidden datasets) to increase the odds of a wrongful conviction based on 911 call audio. Deceptio AI's website is excluded from the Internet Archive; this is a clear indication that Deceptio AI believes itself better off hidden.

This practice is spreading throughout the law enforcement system faster than judges and investigators grounded in reality can possibly counter it.

The emergence of a webpage where a LEO can anonymously upload an audio clip and receive a "guilty" or "not guilty" certificate will crystallize the cost of this issue.

From [3]:

A couple of years ago, he and his two business partners, including one who has decades of experience in statement analysis, decided to join forces and create software that essentially has the brain of a veteran analyst.

“We've come up with an AI now that can detect deception in a person's written or spoken words,” Carson said.

In simple terms, Carson said a client of the business would go on their website, Deceptio.AI, and companies that purchase the software can input statements and the program determines how truthful the statement is and why it may or may not be the whole truth.

“Then we're going to simply click analyze statement and then what the section does is it gives you a probability of truthfulness,” Carson said when demonstrating how Deceptio works. “Now, what we see is anything that falls 85% and under means it's a highly deceptive statement.”

From [4]:

He designed the platform for widespread usage and said it requires no training. The co-founders have bootstrapped the startup over the past two years and recently opened their first funding round.

“We’re seeking the right investors,” Carson explained. “Those are the ones that understand the goal and vision of what a true AI tool like this will mean long-term. We’ve turned down a couple already.”

Carson said the company has also collected and stored a massive amount of human behavioral data, called pattern of life analysis. He said Deceptio’s database “literally maps” deceptiveness in the human psyche.

He noted that Cathie Wood, CEO of St. Petersburg-based ARK Invest, frequently mentions the value of AI entrepreneurs amassing proprietary data. Carson called Deceptio’s information, which does not include personal information, “exceptionally proprietary.”

“To our knowledge, there isn’t anyone else on the planet doing what we’re doing,” he added. “Let alone amassing the type of life intelligence data we’re collecting.”

[1] Statement Analysis by Mark McClish and Tracy Harpster: https://web.archive.org/web/20231004064240/https://www.statementanalysis.com/bio/
[2] Deceptio AI: https://www.deceptio.ai/
[3] https://web.archive.org/web/20240325092556/https://baynews9.com/fl/tampa/news/2023/09/14/deceptio-ai-detects-lies
[4] https://web.archive.org/web/20231002083318/https://stpetecatalyst.com/st-pete-startup-uses-ai-to-detect-lies/

If you believe similarly useless discrimators and their corporate reproductive organs will not be created for other abusive purposes, e.g. by banks, landlords, insurance firms, school administrators, university regents, forensic investigators, farmers, miners, doctors, or nurses, you are simply not paying attention.

  • Pytorch version: 2.2.1
  • Operating System and version: Ubuntu 20.04

Your Environment

  • Installed using source? [yes/no]: yes
  • Are you planning to deploy it using docker container? [yes/no]: no
  • Is it a CPU or GPU environment?: Both
  • Which example are you using: all
  • Link to code or data to repro [if any]:

Expected Behavior

PyTorch should prohibit users from creating discriminators or generators intended for use on the real world which are trained with data not representative of the real world.

Current Behavior

Anyone with an NVIDIA GPU can download PyTorch and train a model on fake datasets, then re-sell access to the model as an "investigative service."

Possible Solution

Destroy PyTorch.

Steps to Reproduce

Deceptio.AI

  1. https://www.propublica.org/article/911-call-analysis-fbi-police-courts
  2. https://www.propublica.org/article/911-call-analysis-jessica-logan-evidence

Versions

Collecting environment information...
PyTorch version: N/A
Is debug build: N/A
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: N/A
OS: Ubuntu Linux 22.04 LTS