Big Technology Podcast Summary: AI’s Drawbacks: Environmental Damage, Bad Benchmarks, Outsourcing Thinking

Podcast Information:

• Title: Big Technology Podcast
• Host: Alex Kantrowitz
• Episode: AI’s Drawbacks: Environmental Damage, Bad Benchmarks, Outsourcing Thinking — With Emily M. Bender and Alex Hanna
• Release Date: May 14, 2025

Introduction

In this illuminating episode of the Big Technology Podcast, host Alex Kantrowitz engages in a thoughtful discussion with two prominent critics of artificial intelligence (AI): Professor Emily M. Bender, a linguistics scholar from the University of Washington, and Alex Hanna, Director of Research at the Distributed AI Research Institute. The conversation delves into the multifaceted drawbacks of AI, including its environmental impact, the reliability of performance benchmarks, and the implications of outsourcing cognitive tasks to machines.

Environmental Impact of AI

Leah Smart initiates the conversation by addressing the significant energy consumption associated with large language models (LLMs). Both guests provide insights into the hidden environmental costs of AI technologies.

• Emily M. Bender describes AI as "a parlor trick" that simulates understanding, yet hides substantial environmental consequences. She emphasizes the excessive energy usage:

  “If you're getting back an AI overview which happens non-consensually when you try Google searches these days. Each of those tokens has to be calculated individually and so it's coming out one word at a time and that is far more expensive.”
  (06:30)

• Alex Hanna sheds light on real-world impacts, citing the Memphis data center operated by XAI, which consumes approximately one million gallons of water daily for cooling purposes. He notes:

  “We have to imagine the problems probably exacerbated right now.”
  (08:07)

• Emily further contextualizes the environmental issues by referencing Dr. Sasha Lucioni’s work on the broader impacts of AI, such as communities losing access to vital water sources and the destabilization of electrical grids. She highlights the concealed nature of these effects:

  “The compute and its environmental footprint and the noise and everything else is hidden from you in the immateriality of the cloud.”
  (08:59)

Benchmark Gaming and Validity Issues

The conversation transitions to the reliability of AI performance benchmarks, a topic both critics argue undermines the true capabilities of AI models.

• Emily critiques the construct validity of many AI benchmarks, stating:

  “Most of the benchmarks that are out there are not reasonable. They lack what's called construct validity.”
  (15:04)
  She explains that benchmarks often fail to measure what they claim to, serving more as marketing tools than genuine assessments of AI performance.

• Alex adds depth by discussing specific examples like Med-PaLM and the ARC AGI test, questioning their real-world applicability and the lack of external validation:

  “There's so much guardedness about what's actually happening here. We can't quantify it.”
  (05:29)

• Emily further elaborates on the ARC AGI test, emphasizing its misnomer and limited scope:

  “That's called ARC AGI, that suggests that it's testing for AGI. It's not.”
  (18:15)

Usefulness and Misuse of AI Models

Addressing the practical applications of AI, Leah shares a personal use case where she leverages ChatGPT to plan her activities in Paris by synthesizing information from uploaded documents.

• Leah advocates for AI’s efficiency in organizing information:

  “It's able to go and comb the Internet for these events and then take into context some of the context that I've given it with these documents, I think is very impressive.”
  (27:00)

• Emily counters by highlighting the limitations and potential pitfalls:

  “What ChatGPT can do is it can mimic human language use across many different domains. It can produce the form of a poem... It is an extremely bad idea to use it if you actually have an information need.”
  (24:42)
  She argues that relying on AI for information synthesis can lead to misinformation and distract from engaging with reliable community resources.

AI in the Medical Context

A significant portion of the discussion focuses on the application of AI in healthcare, particularly in medical transcription and administration.

• Leah proposes using AI to transcribe and summarize doctor-patient interactions to reduce the administrative burden on physicians, thereby allowing more time for patient care.

• Emily raises multiple concerns:

  “Writing the clinical note is actually part of the process of care. It is the doctor reflecting on what came out of that conversation...”
  (39:33)
  She emphasizes privacy issues with ambient listening devices and the unequal effectiveness of transcription for diverse patient demographics.

• Alex adds that current AI transcription tools, like OpenAI’s Whisper, often produce inaccurate and harmful outputs, necessitating additional verification:

  “These things include racial commentary, violent rhetoric, and even imagined medical treatments.”
  (41:48)

• Emily underscores the systemic issues, noting that automating transcription does not address the root problem of excessive paperwork driven by insurance requirements:

  “The insurance companies are not providing any value. They are just vampires on our healthcare.”
  (48:40)

Bottom-Up Adoption and Workplace Implications

The discussion explores whether AI tools being adopted organically by workers could mitigate some of the negative impacts or if they would exacerbate existing issues.

• Emily expresses skepticism about bottom-up adoption, suggesting that even when workers find utility in AI tools, it often leads to increased pressure and diminished quality of work:

  “Every time someone says, Well, I need ChatGPT for this, usually one of two things is going on...”
  (54:54)

• Alex warns that organic use can lead to unrealistic expectations and overreliance, potentially marginalizing those who resist AI adoption:

  “Where does that leave the people who are resistors or thinking about, well, I know this can't do a good job, so where's that putting me?”
  (56:15)

Doomerism and AI Safety

In the final segment, Leah challenges the authors' dismissal of "doomerism," the belief that AI poses existential threats.

• Emily defends her stance by differentiating between speculative long-term risks and immediate, tangible harms:

  “Anytime we are taking the focus away, it's like, has that happened? This is still people writing science fiction...”
  (58:28)

• Alex concurs, acknowledging real issues like social media’s negative impacts but critiques the focus on exaggerated future scenarios:

  “There are problems with social media for sure.”
  (59:35)

• Leah counters by emphasizing the importance of preemptive measures to address potential AI threats, drawing parallels with the delayed recognition of social media issues:

  “You don't agree? Yeah, go ahead, Alex.”
  (59:55)

Conclusion

The episode concludes with Leah Smart summarizing the key points and encouraging listeners to engage critically with AI technologies. She underscores the importance of understanding both the benefits and drawbacks presented by AI, urging a balanced approach to its integration into various sectors.

“For those who are listening or watching, and you may not agree with everything, either everything I said or everything our guests said, hey, at least now you know these arguments and you know the arguments for and against and we trust you to make up your own opinion and do further research.”
(61:06)

Emily M. Bender and Alex Hanna express their gratitude, with Emily wishing Leah a pleasant time in Paris and Alex mentioning Clara, his cat, humorously highlighting the personal aspects behind their critical stances.

Key Takeaways

1. Environmental Concerns: AI’s substantial energy consumption and resource usage have significant, often underreported environmental impacts.

2. Benchmark Reliability: Current AI benchmarks are frequently flawed, lacking construct and external validity, which undermines their usefulness in accurately assessing AI capabilities.

3. Practical Usefulness: While AI tools like ChatGPT offer efficiency in information synthesis, they pose risks related to accuracy, misinformation, and reduced human engagement with reliable information sources.

4. AI in Healthcare: Automating medical transcription with AI may exacerbate existing systemic issues, compromise patient-doctor interactions, and introduce privacy and accuracy concerns.

5. Workplace Dynamics: Bottom-up adoption of AI tools by workers can lead to increased pressure, overreliance, and marginalization of those who choose not to use AI.

6. Doomerism vs. Real-World Harms: While speculative long-term AI risks receive significant attention, immediate and tangible harms, such as environmental damage and systemic inefficiencies, require focused mitigation efforts.

For further reading:

• Book Mentioned: AIcon: How to Fight Big Tech's Hype and Create the Future We Want by Emily M. Bender and Alex Hanna.
• Related Works:
  • Stochastic Parrots by Emily M. Bender et al.
  • More Everything Forever by Adam Becker.
  • The Tesseral Bundle of Ideologies by Timnit Gebru and Emil Torres.

Note: This summary is based on the transcript provided and aims to encapsulate the core discussions and viewpoints expressed by the guests during the podcast episode.