Blogs

A company cuts workers, keeps spending on artificial intelligence (AI), and the public explanation collapses into one sentence: AI caused the layoff.

That can be true in some roles. It can also be the wrong causal story. On April 23, 2026, the Associated Press reported that Meta planned to cut about 8,000 workers while continuing to spend heavily on AI infrastructure and AI talent [1]. That is a real business event. It is not, by itself, a clean causal estimate.

A layoff can be associated with AI investment while still being partly driven by market pressure, investor expectations, prior hiring, and competitive repositioning.

Share: The causal question is not whether artificial intelligence appeared in the layoff story, but what changed when the artificial intelligence investment was pushed. #Causality #ArtificialIntelligence #Economy

In this article, we turn the headline into a smaller question:

• how much layoff pressure is caused by pushing AI investment, after separating the background pressures that made both the investment and the layoff more likely?

Why the easy answer is attractive

The easy answer has a clean timeline:

• the company invests in AI
• the company cuts workers
• therefore AI caused the cuts

That story feels natural because the treatment and outcome are both visible. The problem is that the same business environment can push both variables at once.

If investors want lower costs, competitors are racing toward AI, and margins are under pressure, then AI investment is not floating in space. It is selected by the same strategic pressure that can also produce layoffs.

Draw the graph

The directed acyclic graph (DAG), a graph whose arrows do not loop back on themselves, separates the story into parts:

• AIInvestment: how hard management pushes AI infrastructure and AI hiring
• AutomationCapacity: work that can plausibly be automated
• CostPressure: budget pressure from expenses, margins, and spending plans
• LayoffPressure: pressure to cut headcount
• MarketPressure and AICompetition: background causes that can push several nodes at once

Here is the same graph as a py-scm setup, including the toy coefficients. The coefficients are the parameters that make the example executable; they are not measured company estimates.

import numpy as np

from pyscm.reasoning import create_reasoning_model

nodes = [
    "MarketPressure",
    "AICompetition",
    "AIInvestment",
    "AutomationCapacity",
    "CostPressure",
    "LayoffPressure",
]

weighted_edges = [
    ("MarketPressure", "AIInvestment", 0.30),
    ("MarketPressure", "CostPressure", 0.90),
    ("MarketPressure", "LayoffPressure", 0.40),
    ("AICompetition", "AIInvestment", 0.80),
    ("AICompetition", "LayoffPressure", 0.20),
    ("AIInvestment", "AutomationCapacity", 0.60),
    ("AIInvestment", "CostPressure", 0.70),
    ("AIInvestment", "LayoffPressure", 0.20),
    ("AutomationCapacity", "LayoffPressure", 0.50),
    ("CostPressure", "LayoffPressure", 1.10),
]

idx = {node: i for i, node in enumerate(nodes)}
B = np.zeros((len(nodes), len(nodes)))
for parent, child, weight in weighted_edges:
    B[idx[child], idx[parent]] = weight

A = np.eye(len(nodes)) - B
cov = np.linalg.inv(A) @ np.eye(len(nodes)) @ np.linalg.inv(A).T

model = create_reasoning_model(
    {"nodes": nodes, "edges": [(p, c) for p, c, _ in weighted_edges]},
    {"v": nodes, "m": [0.0] * len(nodes), "S": cov.tolist()},
)

The key backdoor path is AIInvestment <- AICompetition -> LayoffPressure. If you do not block that path, the AI variable picks up part of the broader competitive shock.

Build the toy structural model

The toy model is a structural causal model (SCM). In an SCM, each variable is generated from its direct causes. The numbers below are illustrative; they are not estimates for Meta or any other company.

The inference code separates an observed slice from an intervention:

mean, _ = model.pquery({"AIInvestment": 1.0})
raw_slice = float(mean["LayoffPressure"])

do_high = model.iquery("LayoffPressure", {"AIInvestment": 1.0})
do_low = model.iquery("LayoffPressure", {"AIInvestment": 0.0})
intervention = model.equery(
    "LayoffPressure",
    {"AIInvestment": 1.0},
    {"AIInvestment": 0.0},
)

raw_slice asks what layoff pressure looks like when we merely observe high AI investment. do_high and do_low ask what changes when we set AIInvestment directly, and intervention reports the intervention contrast.

Observed layoff-pressure gap: +1.60
Intervention effect:          +1.27
Observed minus intervention:  +0.33

The raw comparison is larger because it carries some of the market and competition pressure that selected the AI investment in the first place.

The intervention effect is still positive in this toy world. The point is not that AI cannot cause layoffs. The point is that the headline comparison is not the same as the causal effect.

What the model makes explicit

The useful output is not just a number. It is the split between three claims:

• pquery style reasoning: layoffs are higher in the observed high-AI-investment slice
• iquery style reasoning: the AI investment intervention still increases layoff pressure
• graph inspection: part of the observed association comes from background pressure

That distinction matters for economic arguments. A manager can cite AI in the memo. Investors can reward efficiency. A company can redirect spending into data centers. A causal model should force those mechanisms into the same graph before anyone says “AI caused the layoff.”

The practical read

A defensible article, dashboard, or policy argument should avoid the one-line causal jump. Ask:

• what jobs were actually automated?
• what spending was reallocated?
• what market pressure existed before the AI plan?
• what comparable teams or firms did not receive the same AI push?

If those questions are not answered, the safest conclusion is narrower:

AI may have been part of the layoff mechanism, but the observed layoff story also contains cost pressure, competition, and strategic selection.

Sources

1. Meta cuts 8,000 jobs as AI spending surges and efficiency drive accelerates, Associated Press, accessed April 28, 2026.

Download the runnable standalone Python example: Python example ZIP.