% ============================================================================
\section{Robustness and Identification}
\label{sec:robustness}
% ============================================================================

The post-2022 change is large on average and, as Section~\ref{subsec:break} showed, concentrated.
This section confronts it with the alternative explanations a sceptical reader would raise: that it
is mechanical, that it is a mega-cap rally in disguise, that it is sensitive to how the signal is
built or to which filings were scored, that it reflects the composition of a current-membership
universe, that it merely tracks the technology sector, and that the inference behind the event
study and the difference-in-differences estimates is overstated. Some of these explanations are
rejected and some are not.

\subsection{Placebo: is the change specific to AI? (H3)}
If the post-2022 spread were a mechanical artefact of the pipeline, or a generic ``high-disclosure
firms did well'' phenomenon, it should appear for any semantic factor. Table~\ref{tab:placebo} and
Figure~\ref{fig:placebo} re-run the sub-period and interaction design on all eight factors. The
three AI factors show post-2022 spreads of $+13.6$ to $+14.3$ points. Macro-risk language shows no
spread ($-0.5$). ESG-commitment and climate-risk language earn negative post-2022 spreads ($-8.8$
and $-5.0$ points), in line with the post-2021 reversal in the sustainable-investing literature
\citep{pastor2022dissecting}. Two non-AI factors, however, also rise: cybersecurity ($+13.5$),
which is bundled with technology exposure and whose anchor itself mentions AI-powered threat
detection, and automation/labour ($+7.2$), whose anchor refers to AI and robotics. The placebo
therefore rules out a generic disclosure effect and a pipeline artefact, and it shows a rotation
away from sustainability language, but it does not isolate AI from the broader technology theme
with which AI, cybersecurity, and automation language travel. In the interaction regressions none
of the eight coefficients other than AI infrastructure is significant, and the three post-period
annual spreads do not support finer distinctions.

\input{tables/tab_placebo}

\begin{figure}[htbp]\centering
\includegraphics[width=0.92\textwidth]{fig7_placebo.png}
\caption{\textbf{Placebo test across semantic factors.} Post-ChatGPT high-minus-low (Q5$-$Q1)
return spread for each of the eight semantic factors. The AI-related factors (in red) and
cybersecurity cluster at the top; automation/labour language also earns a positive spread;
ESG-commitment and climate-risk language earn negative spreads, and macro-risk language none. The
contrast rules out a generic disclosure effect but not a broader technology theme.}\label{fig:placebo}
\end{figure}

\subsection{Event study around ChatGPT}
The portfolio results use annual formation and forward returns; the event study provides a daily,
cleanly dated complement. Table~\ref{tab:event} and Figure~\ref{fig:event} report market-model
cumulative abnormal returns by pre-event AI tercile. The ordering is monotone at every horizon and
widens over time. The High-minus-Low CAR difference is $+2.9$ points at 60 trading days, $+10.1$
points at one year, and $+14.6$ points at eighteen months; the Welch $t$-statistics of the first
version, 1.77, 1.97, and 1.89, treat the 160 High and 162 Low firms as independent. Because
abnormal returns are purged of market beta, the spread is not a difference in beta exposure. A
cross-sectional regression of one-year CAR on pre-event exposure yields $+3.1$ points per standard
deviation \tstat{1.39}; adding sector fixed effects reduces it to $+2.1$ points \tstat{0.99},
again locating much of the effect in a sector channel. Figure~\ref{fig:eventbin} shows that the
relation is smooth across the exposure distribution.

\input{tables/tab_event}

Table~\ref{tab:eventcda} re-examines the inference. The average pairwise correlation of
estimation-window residuals across the 322 High and Low firms is 0.039; for a one-sample test on
$n$ firms, \citet{kolari2010} show that the independence-based $t$-statistic overstates the true
one by the factor $\sqrt{(1+(n-1)\bar\rho)/(1-\bar\rho)}$, which here is almost four. The portfolio
statistic of equation~\eqref{eq:cda} avoids the problem by treating the daily High-minus-Low
abnormal return as one series. With the estimation-window standard deviation it gives $t=1.35$ at
60 days, $1.45$ at six months, $2.28$ at one year, and $2.63$ at eighteen months; with the
event-window standard deviation, $1.54$, $1.65$, $2.60$, and $2.99$. The one-year and
eighteen-month differences are therefore significant under tests that allow for dependence,
whereas the sixty-day difference---the horizon at which an attention shock should be most
visible---is not. The pattern of a difference that accumulates over a year and a half, rather than
appearing within weeks of the launch as in \citet{eisfeldt2023generative}, is more naturally read
as the gradual sector-level repricing that the annual results also display than as a discrete
attention shock, and at such horizons the market-model benchmark is coarse \citep{fama1998}.

\input{tables/tab_event_cda}

\begin{figure}[htbp]\centering
\includegraphics[width=\textwidth]{fig8_event_car.png}
\caption{\textbf{Cumulative abnormal returns around ChatGPT.} The left panel plots the mean
market-model cumulative abnormal return (CAR) by pre-event AI tercile over the event window; the
right panel plots the High$-$Low CAR spread. The tercile ordering is monotone and the spread
widens gradually, reaching about $+10$ percentage points one year after the launch. All terciles
drift negative because the market model was estimated in the 2022 bear market; only differences
between terciles are interpretable.}\label{fig:event}
\end{figure}

\begin{figure}[htbp]\centering
\includegraphics[width=0.66\textwidth]{fig13_event_binscatter.png}
\caption{\textbf{One-year abnormal return versus pre-event AI exposure.} Binned scatter, with
fitted line, of each firm's one-year cumulative abnormal return after ChatGPT against its pre-event
AI exposure (filings of calendar 2021). The relation is positive and smooth across the exposure
distribution.}\label{fig:eventbin}
\end{figure}

\subsection{Difference-in-differences and the cross-sectional nature of the effect (H4)}
Table~\ref{tab:did} reports the difference-in-differences estimates of equation~\eqref{eq:did}, in
which treatment is the firm's pre-determined (2019--2021) exposure interacted with the post-2022
indicator, within firm and year fixed effects. The continuous-exposure coefficient is $+2.7$
percentage points per standard deviation \tstat{1.51}; the binary version is $+1.7$ points
\tstat{0.78}. Both are positive and neither is significant at conventional levels. The
disappearance of the effect under firm fixed effects in Table~\ref{tab:panel} and the weakness of
the DiD estimates say the same thing: once the permanent component of AI exposure is absorbed,
little remains to be priced.

\input{tables/tab_did}

The event-time profile of equation~\eqref{eq:evt}, in Figure~\ref{fig:eventtime}, is less
favourable to the regime interpretation than the first version of this paper reported. With
standard errors clustered by firm, the premium on fixed pre-launch exposure is $+3.9$ points per
standard deviation in 2015 \tstat{2.8}, $+3.0$ in 2016 \tstat{2.3}, and $+2.3$ in 2018
\tstat{1.5}, falls to zero or below in 2017, 2019, and 2020, and is $+3.4$ in 2022 \tstat{2.2},
$+0.5$ in 2023 \tstat{0.3}, and $+6.8$ in 2024 \tstat{3.0}. The 2015 premium is as large as the
2022 premium; only 2024 stands apart. The parallel-trends assumption of the DiD design is thus not
satisfied in the strict sense, and the figure is better read as showing that firms with persistently
AI-intensive filings earned a premium in several years scattered across the sample, with an
unusually large one in the last. The first version's confidence bands, computed with year-clustered
errors, were spuriously narrow (standard errors of 0.1--0.4 points instead of 1.0--2.3), and the
statement that the profile was ``flat and near zero through 2021'' is withdrawn.

\begin{figure}[htbp]\centering
\includegraphics[width=0.82\textwidth]{fig11_eventtime.png}
\caption{\textbf{Event-time profile of the premium on pre-determined AI exposure.} Estimated return
premium on firms' pre-determined (2019--2021) AI exposure by formation year, relative to 2021,
with 95\% confidence intervals from standard errors clustered by firm (equation~\eqref{eq:evt}). The
premium is significantly positive in 2015, 2016, 2022, and 2024, and largest in 2024; the pre-launch
profile is not flat.}
\label{fig:eventtime}
\end{figure}

\subsection{Construction, horizon, and universe robustness}
Table~\ref{tab:robust} collects the construction checks. The post-2022 spread is essentially
unchanged whether the signal is the section-weighted mean (base, $+14.6$), the simple mean
($+13.4$), or the top-5\% tail ($+14.9$); for 8-K filings the first two coincide, so this check
speaks mainly to the 10-Q filings of the 84 firms. The spread survives replacing the three AI
signals with their first principal component, which explains 86.8\% of their common variance
(post-2022 spread $+13.4$). It is $+14.1$ points when the Magnificent~7 are excluded and $+14.5$
when the ten largest firms by market capitalization are dropped (Figure~\ref{fig:mag7}): the
post-period spread is not the footprint of a handful of trillion-dollar winners, which is
unsurprising given that its largest contributors are mid-sized memory and storage firms. Two checks
are less favourable. The six-month forward-return spread in the post-period is $+1.1$ points,
computed on the 2022 and 2023 cohorts only, so the post-2022 relation is not visible at a
six-month horizon in the cohorts that can be evaluated. And the sector-neutral composite, which
compares firms only to others within their GICS sector, produces a post-2022 spread of $+3.8$
points \tstat{0.84}, a quarter of the baseline: most of the spread operates between sectors.

\input{tables/tab_robust}

\begin{figure}[htbp]\centering
\includegraphics[width=0.78\textwidth]{fig10_mag7.png}
\caption{\textbf{The post-2022 spread is not a mega-cap artefact.} Post-2022 composite Q5$-$Q1
return spread for the full index, after excluding the Magnificent~7, and after excluding the ten
largest firms by market capitalization, together with the pre-period spread excluding the
Magnificent~7. The spread is essentially unchanged by the exclusions.}\label{fig:mag7}
\end{figure}

\subsection{Composition of the corpus: the 8-K-only signal}
\label{subsec:corpus}
Because the firm-year score of equation~\eqref{eq:annual} weights filings equally, and because the
10-Q filings that were scored belong to 84 alphabetically contiguous firms and score about 0.07
higher on the AI-adoption anchor than 8-K filings, the baseline signal partly sorts firms on whether
their quarterly reports happened to be parsed. The share of firm-years containing a scored 10-Q
rises from 3\% in the bottom AI quintile to 30\% in the top quintile, and within a formation year a
scored 10-Q raises the composite $z$-score by 0.67 standard deviations \tstat{6.7}. This is a
measurement artefact, not an economic one. Panel~A of Table~\ref{tab:reviewrobust} rebuilds the
composite from 8-K filings only, using the filing-level scores exported from the database
(\texttt{data/processed/filing\_type\_scores.parquet}). The 8-K-only composite has a correlation of
0.95 with the baseline, a scored 10-Q no longer predicts it ($-0.10$ standard deviations,
$t=-1.1$), and its spreads are $+1.6$ points before 2022 and $+16.3$ after \tstat{2.10}, with
cohort values of $+6.3$, $+1.3$, and $+41.3$; its monthly long--short earns a post-period Sharpe
ratio of 1.42 \tstat{2.25}. Dropping the 84 firms altogether gives $+0.6$ before and $+15.0$
after. The post-2022 change is therefore not an artefact of the partially parsed 10-Q sample,
and the paper's results should be read as results about the language of current reports. That
the 2023 cohort turns from $-4.0$ to $+1.3$ when the 10-Q firms are removed from the signal shows,
however, that cohort-level spreads are sensitive to measurement details of this kind.

\input{tables/tab_review_robust}

\subsection{Entrants and survivorship}
\label{subsec:entrants}
Panel~B of Table~\ref{tab:reviewrobust} addresses the current-membership universe. Firms that
joined the index after 2010 and have no price history before July 2009 earn higher average returns
than seasoned firms (they were selected into the index partly because of those returns), but they
are spread evenly across the AI quintiles (11--12\% of each), so they do not generate the
cross-sectional pattern. Among the 416 seasoned firms the composite spread is $+1.0$ point before
2022 and $+17.3$ after \tstat{2.36}, with cohort values of $+9.1$, $+1.4$, and $+41.3$; among
firm-years with at least 36 months of price history at formation it is $+2.8$ and $+14.9$. The
check does not address the firms that left the index before 2026, for which we have no data; the
level of returns is biased upward by their absence, and a firm that disclosed heavily on AI and
subsequently failed is missing from the bottom of the return distribution.

\subsection{Definition of the break}
Panel~C of Table~\ref{tab:reviewrobust} moves the 2021 cohort, whose holding period straddles the
launch. Ending the pre-period in 2020 gives a pre-period spread of $+2.0$ points \tstat{1.17};
starting the post-period in 2021 gives $+12.7$ points over four cohorts, of which the first three
average $+4.1$. The conclusion that the post-period average depends on the 2024 cohort is
unaffected by where the line is drawn.

\subsection{Sector controls and the technology factor}
\label{subsec:sector}
Panel~D of Table~\ref{tab:reviewrobust} and Panel~C of Table~\ref{tab:postinf} address the concern
that AI language proxies for the technology sector. In annual cross-sectional regressions with
size, momentum, and volatility controls, the post-period slope on the composite is $+5.5$ points
per standard deviation \tstat{2.00}; with GICS-sector dummies it is $+2.9$ \tstat{1.36}, and the
2024 slope falls from $+14.3$ to $+10.2$. In pooled regressions of the winsorized forward return
with firm-clustered errors, year fixed effects give $+2.9$ \tstat{2.76} and sector-by-year fixed
effects $+1.3$ \tstat{1.22}. In the monthly series the long--short loads 0.47 on the
technology-sector factor after 2022 and its FF6 alpha falls from $+5.2\%$ to $+1.9\%$ per year
\tstat{0.86} when the factor is added; the pre-period FF6 alpha with the sector factor is
$-0.4\%$. A within-sector premium on AI language after 2022 therefore exists in point estimate but
is not statistically distinguishable from zero in any of these specifications; the statistically
detectable part of the post-2022 change is the between-sector part.

\subsection{An alternative performance outcome: risk}
Performance is not only about mean returns. High-AI firms exhibit modestly higher forward realized
volatility, and the gap widens after 2022 (a Q5$-$Q1 difference of about 2.9 annualized volatility
points, $t=3.1$, against 2.0 points before). Part of the post-2022 return differential is
compensation for the greater volatility of AI-exposed firms during a period of rapid technological
repricing.